Episode Transcript
Transcripts are displayed as originally observed. Some content, including advertisements may have changed.
Use Ctrl + F to search
10:00
have genius, they actually listen, get inspired,
10:02
and then they come out with something
10:04
different, something new. They don't blend around
10:06
patterns based on machine-based algorithms. So nice
10:09
try, but I don't think that that
10:11
argument is very convincing. And I also
10:13
love that they say that, you know,
10:16
the creators and their partners are the
10:18
ones that have resorted to like the
10:20
old legal playbook. They're
10:23
not resorting to, oh, we can
10:25
do this. It's based on fair use. It's
10:27
transformative. We're going to seek forgiveness
10:29
instead of permission. Well, I mean, you also have
10:31
the investor in the company who you quote in
10:34
the lawsuit saying, because he said this to a
10:36
news outlet, I don't know if I would have
10:38
invested in this company if he had a deal
10:40
with the record labels, because then they probably wouldn't
10:42
have needed to do what they needed to do,
10:44
which I assume he sort of meant, you know,
10:47
hoover up all this music without paying for it.
10:50
Yeah, that's in the legal world what we call a
10:52
bad fact. That is a
10:54
bad fact for the other side. You
10:57
don't want your investors saying, gee, you know, if
10:59
they had really done this the legal way, I
11:01
don't think I would have invested, because it's just
11:03
too hard. It's just too hard to do it
11:05
the legal way. Mitch, we've
11:07
seen other lawsuits come out in the past
11:10
year from media companies, including
11:12
the New York Times, which sued
11:14
OpenAI and Microsoft last year, alleging
11:17
similar types of copyright violations. How
11:19
similar or different from the sort
11:21
of text-based copyright arguments is
11:23
the argument that you are making against
11:26
these AI music generation companies? I
11:28
think the arguments are the same that
11:31
you have to get permission before you copy.
11:33
It's just basic copyright law. The
11:35
businesses are very different. And I think
11:38
looking at sort of the
11:40
public reports on the licensing negotiations
11:42
going on between the news media
11:44
and companies like
11:46
OpenAI, where news is
11:48
dynamic. It has to change every single day.
11:51
And so there needs to be a
11:53
feed every single day for the input
11:55
to actually be useful for the output.
11:58
Music is catalyzing. a
18:00
smile doing our job. Here,
18:03
I think that really
18:05
what we're trying to do is
18:08
create a marketplace like streaming,
18:10
where there are partnerships and
18:13
both sides can grow and evolve together. Because the truth
18:15
is you don't have one without the other. Record
18:19
companies don't control their prices, they
18:21
don't control their distribution. They're now
18:23
gateways, not gatekeepers. The democratization of
18:25
the music industry has changed everything.
18:28
And I think they're seeking the same
18:30
kind of relationships with AI companies that
18:32
they have with streaming companies today. What
18:35
would a good model look like? I mean, there
18:37
are reports this week that YouTube is in talks
18:39
with record labels about paying them a lot of
18:41
money to license songs for their
18:43
AI music generation software. Do you think
18:45
that's the solution here? That there will
18:47
be sort of these platforms that pay
18:49
record labels and then they get to
18:52
use those labels, songs in
18:54
training their models. Do you
18:56
think it's fine to use AI to generate music
18:58
as long as the labels get paid? Or is
19:00
there sort of a larger objection to the way
19:02
that these models work at all? I
19:05
think it works as long as it's done
19:07
in partnership with the artists. And at the
19:09
end of the day, it moves
19:11
the ball forward for the label and the
19:14
artist. I mean, the YouTube example
19:17
is interesting because that's really geared
19:20
towards YouTube shorts, right? It's
19:22
geared towards fans being
19:24
able to use generated music
19:26
to put with their own videos for 15
19:29
or 30 seconds. That's
19:31
an interesting business model. BandLab
19:33
is a tool for artists, splice,
19:36
beatport, focus, right,
19:39
output, waves, even tide.
19:42
Every digital audio workstation that's now
19:44
using AI, native instruments, Oberheim. I
19:47
mean, there are so many AI
19:50
companies that have these bespoke agreements
19:52
and different types of tools that
19:54
are meant to be done with
19:57
the artistic community that
19:59
I think they are really important.
20:01
The outliers are the Sunos and
20:03
the Yudios, who frankly are not
20:05
very creative in trying to help
20:08
with human ingenuity. Instead, they're just
20:10
substitutional to make money for investors
20:12
by taking everybody else's stuff. We've
20:15
seen some pretty different reactions to
20:18
the rise of AI among artists.
20:20
Some people clearly seem to
20:22
want no part of it. On the other
20:24
hand, we've seen musicians like Grimes saying, here,
20:26
take my voice, make whatever you want. We'll
20:29
figure out a way to share the royalties
20:31
if any of your songs becomes ahead. I'm
20:33
curious, if you're able to get the deals
20:35
that you want, do you expect any controversy
20:37
within the artist community and artists saying, hey,
20:39
why'd you sell my back catalog to this
20:41
blender? I don't want to be part of
20:44
that. Yeah, I think, look, artists are entitled to be
20:46
different and there are going to be artists, I think
20:48
Kevin, you said earlier, you know artists who are so
20:50
scared of this, they do want to shut the whole
20:53
thing down. They just don't want their music and
20:55
their art touched, right? I know directors
20:57
of movies who can't stand that the
20:59
formatting is different for an airplane. Like
21:01
that's their baby and they just don't
21:04
want it. Then there are artists like
21:06
Grimes who are like, I'm fine being
21:08
experimental. I'm fine having fans take it
21:10
and change it and do something with
21:12
it. All of that is good. They're
21:14
the artist, right? I mean, it's their art. Our
21:16
job is to invest in them, partner with
21:19
them, help find a market for them. But
21:21
at the end of the day, if you're trying
21:23
to find a market for an artist's work that
21:25
they don't and they don't want that work in
21:27
the market, it's not going to work. Yeah.
21:31
Have you listened to much AI generated music? Are
21:33
there any songs you've heard that you thought that's
21:35
actually kind of good? Yeah.
21:38
So I think
21:40
in the sort of overdubbing voice and likeness thing,
21:43
that it's a little bit better than some
21:46
of the simple prompts
21:48
on these AI generators like Houdeo
21:50
and Suno. But I
21:54
heard Billie Eilish's voice on a
21:56
revivalist song and I was like, wow, she
21:58
should cover this song. from
24:00
the Valley? You
24:02
know, this has been the same argument that the Valley's
24:04
had since 1998. To
24:07
me, that's a 30-year-old argument. If
24:10
you look at the marketplace today,
24:13
where Silicon Valley thrives is when
24:15
rights are in place and they
24:17
form partnerships, and then they grow
24:19
into sophisticated global leaders where
24:21
they can tweak, you know,
24:24
every couple of years, their deals
24:26
and come up with new products
24:28
that allow them to feed these
24:30
devices that are nothing without the
24:32
content on them. And, you know,
24:34
there's always sort of this David
24:36
versus Goliath thing, no matter what
24:38
side you're on. But if you
24:40
think about it, music,
24:43
which is a $17 billion industry in the United
24:45
States. I mean, I don't even, I think one
24:48
tech company's cash on hand is five times that,
24:50
right? Not to mention their $289 billion market caps,
24:52
right? But
24:55
they are completely dependent on
24:58
the music that these geniuses
25:00
create in order to thrive.
25:02
And to say that these
25:04
creators are stopping their progress,
25:07
I think is sort of laughable. I
25:09
think what's much more threatening is
25:11
if you move fast and break
25:13
things without partnerships, what
25:15
are you threatening on the tech side
25:17
with a no holds barred, you know,
25:20
culture destroying, you know, machine led world?
25:22
It sounds pretty gross to me. So
25:25
what happens next? The lawsuits have been filed. This
25:27
stuff tends to take a long time, but what
25:29
can we look forward to? You know,
25:31
will there be sort of scandalous emails unearthed
25:34
in Discovery that you'll post to your website? Or what
25:36
can we look forward to here? Well,
25:39
moving forward in Discovery, I think we'll
25:41
be prohibited from posting any thing to
25:44
it. I know, you think
25:46
you're disappointed. If you want to just
25:48
send them to hardforkatnytimes.com, that's
25:50
fine. I live for that stuff.
25:52
But we will, of course, follow
25:54
the rules. But, you
25:57
know, we have filed in the districts
25:59
where these companies. these reside. And
26:01
so I hope that within a year or
26:03
so we will actually get to the meat
26:06
of this because if you think about it,
26:08
at the the judge has to decide
26:11
when they raise fair use as a defense, is
26:13
this fair use or not? Right.
26:15
And that is something that, you
26:17
know, has to be part
26:20
of the beginning part of the lawsuit. So
26:22
we're hopeful that, you know, when
26:24
I say a short time in legal terms,
26:26
that means you know, a year or two,
26:28
but we're hoping that in a short time,
26:30
we will actually get a decision. And that
26:32
it sends the right message to investors and
26:34
to new companies, like there's a right way
26:36
and a wrong way to do this, doors
26:38
are open for the right way. Yeah, I
26:41
think there's a story here about startups
26:43
that are sort of moving fast, breaking
26:45
things asking for forgiveness, not permission. But
26:48
I also think there's a story here
26:50
that that maybe we haven't talked about
26:52
about restraint, because I know that a
26:54
lot of the big AI companies had
26:57
tools years ago that could generate music,
27:00
but they did not release them. I remember
27:02
hearing a demo from someone who worked at the big
27:04
AI companies, one of the big AI companies, maybe
27:07
two years ago, of one of these kinds
27:09
of tools. But I think they understood they
27:11
were scared because they knew that the record
27:13
industry is very organized, it has this
27:15
kind of like, you know, history of
27:17
litigation. And, you know,
27:19
they sort of understood that they were likely
27:21
to face lawsuits if they let this out
27:23
into the public. So have you had discussions
27:25
with the big the bigger AI companies, the
27:28
more established ones that are working on this
27:30
stuff? Or are they just sort of intuiting
27:32
correctly that they would have a lot of
27:34
legal problems on their hands if they let
27:36
this stuff out into the general public? You
27:39
know, you're raising a point that I don't think
27:41
is discussed often enough, which is that there are
27:43
companies out there that
27:46
deserve credit for restraint. And part
27:48
of it is that they
27:51
know that we would bring a lawsuit and in the past,
27:53
we haven't been shy, and that's useful. But
27:55
part of it is also because these are their partners
27:57
now. You know, there are
28:00
real business relationships here and
28:02
human relationships here between these
28:04
companies. And
28:06
so they, they're natural. I
28:10
think they're moving towards a world where
28:12
their natural instinct is to approach
28:14
their partners and see if they can work with them. You
28:17
know, I know that YouTube did
28:20
sort of its Dreamcast experiment, approached
28:22
artists, approached record companies. That
28:24
was sort of like the precursor or the
28:27
beta to whatever they might be discussing now for
28:29
what's going to go on shorts that we talked
28:31
about earlier. And I'm sure that there are many
28:33
others, but you're right. Yes, there
28:35
are going to be companies like Sunu and
28:37
Yudio that just seek investment,
28:39
want to make a profit and steal stuff.
28:42
But there is, there is
28:44
restraint and constructive action
28:47
by a lot of companies out there who
28:50
do view the creators as
28:52
their partners. Well,
28:54
it's a really interesting development and I
28:56
look forward to following it as it
28:58
progresses. Thanks Mitch.
29:00
Thanks so much, Mitch. Thanks for coming by.
29:02
Thanks guys. Bye. When
29:06
we come back, we're going Inside the Pentagon
29:08
with Chris Kirchhoff, the author of Unit X.
29:11
Are we allowed inside the Pentagon? This
29:35
podcast is supported by KPMG. Your
29:38
task as a visionary leader is simple. Harness
29:41
the power of AI. Shape the future
29:43
of business. Oh, and do
29:45
it before anyone else does without leaving people
29:47
behind or running into unforeseen risks. Simple,
29:50
right? KPMG's got you. Helping
29:53
you lead a people-powered transformation that
29:55
accelerates AI's value with confidence. Close
29:58
that for a vision. Learn more
30:00
at... www.kpmg.us.ai Hi,
30:04
I'm Robert Vinlouin from New York Times Games. I'm
30:06
here talking to people about Wordle and showing them
30:09
this new feature. Do you all play Wordle? Yeah.
30:11
I have something exciting to show you. Oh, okay.
30:13
It's the Wordle Archive. Oh! And
30:16
you keep your... So if I miss it, I can like go back.
30:19
Oh, that's nice. So now you can play
30:21
every Wordle that has ever existed. There's like
30:23
a thousand puzzles. Oh my god, I love
30:25
it. Actually, that's really great. What
30:28
date would you pick? May 17th. Okay. That's
30:30
our birthday. What are some of your like
30:32
habits for playing Wordle? I wake up, I
30:34
make a cup of coffee, I do the
30:37
Wordle, and I send it to my friends in a
30:39
group chat. Amazing. Thanks so much for coming by and
30:41
talking to us and playing. New York
30:43
Times Games subscribers can now access the entire Wordle
30:45
Archive. Find out more at nytimes.com/games.
30:47
You don't understand how much Wordle means
30:49
to us. We need to take a
30:51
selfie. Well,
30:56
Casey, let's talk about war. Let's talk
30:58
about war. And what is it
31:00
good for? Some say
31:03
absolutely nothing. Others write books arguing the
31:05
opposite. Yeah. So
31:08
I've been wanting to talk about AI and
31:11
technology and the military for a while on
31:13
the show now, because
31:15
I think what's really flying under the radar
31:17
of kind of the mainstream tech press these
31:19
days is that there's just
31:21
been a huge shift in Silicon
31:24
Valley toward making things for the
31:26
military and the U.S. military in
31:28
particular. You know, years ago,
31:30
it was the case that most of the
31:32
big tech companies, they were sort of very
31:34
reluctant to work with the military to
31:37
sell things to the Department of Defense
31:39
to make products that could be used
31:41
in war. They had a lot of
31:43
ethical and moral quandaries about that, and
31:45
their employees did too. But we've
31:47
really seen a shift over the past few years.
31:50
There are now a bunch of
31:52
startups working in defense tech, making
31:54
things that are designed to be
31:56
sold to the military and to
31:58
national security forces. And
32:00
we've also just seen a big
32:03
effort at the Pentagon to modernize
32:05
their infrastructure, to update their technology,
32:08
to not get beat by other nations when
32:10
it comes to having the latest and greatest
32:12
weapons. Yeah, and also Kevin, just the rise
32:14
of AI in general, I think has a
32:16
lot of people curious about what the military
32:18
thinks of what is going on out here.
32:20
And is it eventually going to have to
32:22
adopt a much more aggressive AI strategy than
32:24
the one it has today? Yeah, so a
32:27
few weeks ago, I met a guy named
32:29
Chris Kirchhoff. He's one of the authors
32:31
along with Raj Shah of a book called Unit
32:33
X. Chris is sort of
32:35
a long time defense tech guy. He
32:37
was involved in a number of tech
32:39
projects for the military. He worked at
32:41
the National Security Council during the Obama
32:43
administration. Fun fact, he was the highest
32:46
ranking openly gay advisor in the Department
32:48
of Defense for years. And
32:50
most importantly, he was a founding
32:52
partner of something called the Defense
32:55
Innovation Unit or DIU. It also
32:57
goes by the name Unit X,
33:00
which is basically this little experimental division
33:02
that was set up about a decade
33:05
ago by the Department of Defense who
33:07
tried to basically bring the Pentagon's technology
33:09
up to date. And
33:12
he and Raj Shah, who was another founding
33:14
partner of the DIU, just wrote a book
33:17
called Unit X that basically tells the story
33:19
of how the Pentagon sort of realized
33:21
that it had a problem with
33:23
technology and set out to fix
33:26
it. So I just thought we should
33:28
bring in Chris to talk about some of the changes
33:30
that he has seen in the military
33:32
when it comes to technology and in Silicon Valley
33:34
when it comes to the military. Let's
33:36
do it. ["The Military"]
33:50
Chris Kirchoff, welcome to Hard Fork. Glad
33:52
to be here. So I think
33:54
people hear a lot about the military and
33:56
technology and they kind of assume that they're
33:59
like very futuric. things happening inside the
34:01
Pentagon that we'll hear about at some point
34:03
in the future. But a
34:05
lot of what's in your book
34:07
is actually about old technology and
34:09
how underwhelming some of the military's
34:11
technological prowess is. Your book
34:14
opens with an anecdote about your
34:16
co-author actually using a
34:18
compact digital assistant because
34:21
it was better, it had better
34:23
navigation tools than the navigation system
34:25
on his 30 million dollar jet.
34:28
That was sort of how you introduced the
34:30
the fact that the military is not quite
34:32
as technologically sophisticated as many people might think.
34:35
So I'm curious when you first started your
34:37
work with the military, what was the
34:39
state of the technology? Well
34:41
it's it's really interesting to me. You
34:44
go to the movies and we have
34:46
all seen Mission Impossible and James Bond
34:48
and wouldn't it be wonderful if that
34:50
actually were the reality behind the curtain.
34:52
But when you open up the curtain
34:54
you realize that actually in this country
34:56
there are two entirely different systems of
34:58
technological production. There's one for the military
35:00
and then there's one for everything else. And
35:02
to dramatize this on the image of our book
35:04
Unit X we have an iPhone and on top
35:06
of the iPhone is sitting an F-35, the world's
35:10
most advanced fighter jet, a
35:12
fifth-generation stealth fighter known as
35:14
a flying computer for its incredible sensor fusion
35:17
and weapon suites. But the thing about the
35:19
F-35 is that its design was
35:21
actually finalized in 2001 and it did
35:24
not enter operations until 2016.
35:26
And a lot happened
35:29
between 2001 and 2016 including
35:32
the invention of the iPhone, which by
35:34
the way has a faster processor in
35:36
it than the F-35. And
35:38
if you think about the F-35 over
35:40
the subsequent years there's been three technological
35:42
upgrades to it and we're now what
35:44
we're almost an iPhone
35:46
16 season. And once you understand
35:48
that you understand why it was
35:51
really important that the Pentagon thought
35:53
about establishing a Silicon Valley office
35:55
to start accessing this whole other
35:57
technology ecosystem that is faster and
36:00
and generally a lot less expensive than the
36:02
firms that produce technology for the military. Yeah,
36:05
I remember years ago, I interviewed your
36:07
former boss, Ash Carter, the former Secretary
36:09
of Defense who died in 2022. And
36:14
I sort of expected that he'd want
36:16
to talk about like all the newfangled
36:18
stuff that the Pentagon was making, you
36:20
know, autonomous drones, stealth bombers. But
36:22
instead we ended up talking about procurement,
36:24
which is basically how the government buys
36:27
stuff, whether it's a fighter jet or
36:29
an iPhone. And I remember
36:31
him telling me that procurement was
36:33
just unbelievably complicated. And it
36:35
was a huge part of what made
36:38
government and the military in particular so
36:41
inefficient and kind of backwards technologically.
36:44
Describe how the
36:46
military procures things and then what
36:48
you discovered about how to maybe
36:50
short circuit that process or make
36:53
it more efficient. You
36:55
know, if you're looking to buy a nuclear aircraft
36:57
carrier or a nuclear submarine, you can't really go
36:59
on Amazon and price shop for that. There's really
37:01
only- I learned that the hard way, by the
37:03
way. Should have upped your credit, let me case
37:06
you. And so,
37:08
you know, in those circumstances when the
37:10
government is representing the taxpayer and buying
37:13
one large military system, a multi-billion dollar
37:15
system from one vendor, it's
37:17
really important that the taxpayer not be
37:20
overcharged. And so the Pentagon has developed
37:22
a really elaborate system of procurement to
37:24
ensure that it can control how production
37:27
happens, the cost of individual items. And
37:30
that works okay if you're in a situation
37:32
where you have the government and one firm
37:34
that makes one thing. It
37:36
doesn't make any sense though, if you're
37:38
buying goods that multiple firms
37:41
make or that are just available on the
37:43
consumer market. And so one
37:45
of the challenges we had out here in
37:47
Silicon Valley when we first did a Defense
37:49
Innovation Unit was trying to figure out how
37:51
to work with startups and tech companies who
37:53
it turns out weren't interested in working with
37:55
the government. And the reason why is that
37:58
the government typically buys defense. technology
38:00
through something called the federal acquisition
38:02
rules, which is a little
38:04
bit like the Old Testament. It's this
38:07
dictionary size book of regulations. Letting a
38:09
contract takes 18 to 24 months.
38:11
If you're a startup, your investors tell you not
38:13
to go down that path for a couple reasons.
38:16
One, you're not going to make enough money before
38:18
your next valuation. You're going to have to wait
38:20
too long. You're going to go out of business
38:22
before the government actually closes the sale. And
38:24
two, even if you get that first
38:26
contract, it's totally possible another firm with
38:28
better lobbyists is going to take it
38:30
right back away from you. So
38:33
at Defense Innovation Unit, we had to figure out
38:35
how to solve that paradox. Part
38:37
of what I found interesting about your
38:40
book was just the sort of
38:42
accounts that you gave of these
38:44
sort of clever loopholes that you
38:46
and your team found around some
38:49
of the bureaucratic slowness at the
38:51
Pentagon. And in particular, this loophole
38:53
that allowed you to purchase technology
38:55
much, much more quickly than one
38:57
of your staffers found. Tell that
38:59
story and maybe that'll help people
39:01
understand kind of the systems that
39:03
you were up against. It's
39:05
an amazing story, but we knew when we
39:08
arrived in Silicon Valley that we would fail
39:10
unless we figured out a different way to
39:12
contract with firms. And our first week in
39:14
the office, this 29-year-old
39:17
staff member named Lauren Daly, the
39:19
daughter actually of a tank commander
39:22
whose way of serving was to become a civilian in
39:24
the Pentagon and work on acquisition, happened
39:26
to be up because she's a
39:28
total acquisition nerd late at night
39:31
reading the just released National Defense
39:33
Authorization Act, which is another dictionary
39:35
size compendium of law that comes
39:37
out every year. And
39:39
she was flipping through it trying to find
39:41
new provisions in law that might change how
39:43
acquisition worked. And sure enough, in section 815
39:46
of the law, she
39:49
found a single sentence that she realized
39:51
somebody had placed there that changed everything.
39:54
And that single sentence would allow us
39:57
to use a completely different kind of
39:59
contracting mechanisms called other. transaction authorities that
40:01
were actually first invented during the space
40:03
race to allow NASA during the Apollo
40:05
era to contract with mom and pop
40:08
suppliers. And so she realized
40:10
that this provision would allow us
40:12
not only to use OTAs to buy technology,
40:15
but the really important part is that if
40:17
it worked, if it was successful in the
40:19
pilot, we could immediately go to buy it
40:21
at scale, to buy it in production. We
40:23
didn't have to recompete it. There would
40:26
be no pause, no 18-month pause between
40:28
demonstrating your technology and having the department
40:30
buy it. And when Lauren brought
40:32
this to our attention, we thought, oh boy, this really is
40:34
a game changer. So
40:36
we flew Lauren to Washington. We had
40:38
her meet with the head of acquisition policy at the Department of
40:41
Defense. And in
40:43
literally three weeks, we changed 60 years
40:45
of Pentagon policy to create a
40:47
whole new way to buy technology that to this
40:49
day has been used to purchase $70 billion of
40:52
technology for the Department of Defense. You
40:54
just said that the reason that Silicon Valley
40:57
tech companies, some of them didn't want to
40:59
work with the military is because of
41:01
this sort of arcane and complicated
41:04
procurement process. But
41:06
there are also real moral objections among
41:08
a lot of tech companies and tech
41:11
workers. In 2018, Google
41:13
employees famously objected to something
41:15
called Project Maven, which
41:18
was a project the company had planned
41:20
with the Pentagon that would have used
41:22
their AI image recognition software to improve
41:25
weapons and things like that. There
41:27
have been just a lot of objections over
41:29
the years from Silicon Valley to working with
41:32
the military to being defense contractors.
41:35
Why do you think that was and do you think that's changed
41:37
at all? It's
41:40
completely understandable. So few
41:42
Americans serve in uniform. Most of us don't
41:45
actually know somebody who's in the military. And
41:47
it's really easy here in Silicon Valley where
41:50
the weather's great. Sure
41:52
you read headlines in the news, but the military
41:54
is not something that you encounter in your daily
41:56
life. And you join a tech
41:58
company to make the world to develop products
42:00
that are gonna help people. You don't join
42:03
a tech company assuming that you're gonna be
42:05
making the world a more lethal place. But
42:08
at the same time, you know, Project
42:10
Maven was actually something that I got a chance
42:12
to work on and Defense Innovation Unit and a
42:14
whole group of people led. And-
42:17
And tell us about, remind us what Project
42:20
Maven was. So Project Maven was an attempt
42:22
to use artificial intelligence and machine learning to
42:25
take a whole bunch of footage,
42:27
surveillance footage that was being captured
42:29
in places like Iraq and Afghanistan
42:31
and other military missions and to
42:33
use machine learning to label what
42:37
was found in this footage. So
42:39
it was a tool to essentially automate
42:41
work that otherwise would have taken human
42:44
analysts hundreds of hours to do. And
42:46
it was used primarily for intelligence
42:48
and reconnaissance and force protection. So Project
42:51
Maven, this is another misconception. You know,
42:53
when you talk about military systems, there's
42:55
really a lot of unpacking you have
42:57
to do. The headline that
42:59
got Project Maven in trouble said, you
43:02
know, Google working on secret drone project. And
43:04
it made it look as if Google
43:06
was partnering with Defense Innovation Unit, the
43:09
partner of Defense, to build offensive weapons
43:11
to support the US drone campaign. And
43:14
that's not all what was happening. What was
43:16
happening is Google was building tools that
43:18
would help our analysts process the incredible
43:21
amount of data flowing off many
43:23
different observation platforms in the military. Right.
43:26
But Google employees objected to this. They
43:28
made a big case that Google should
43:30
not participate in Project
43:32
Maven and eventually the company pulled
43:34
out of the project. But speaking
43:36
of Project Maven, I was curious
43:38
because there was some reporting from
43:40
Bloomberg this year that showed that
43:42
the military has actually used Project
43:45
Maven's technology as recently as
43:47
February to identify targets for airstrikes in
43:49
the Middle East. So isn't
43:51
that exactly what the Google employees who were protesting
43:53
Project Maven back when you were working on it
43:55
at the Defense Department? Isn't that exactly what they
43:57
were scared would happen? Well, AI
48:00
drone. Am I hearing you
48:02
right that you're saying that we just
48:04
we have to have such overwhelmingly powerful,
48:06
lethal technology in our military that other
48:08
countries won't mess with us? I
48:11
totally hear you and frankly hear all the
48:13
people that you know years ago were flighted
48:15
with the stop killer robots movement.
48:18
I mean these weapons are they're awful
48:20
things. They do awful things to human
48:22
beings. But you know at the
48:24
same time there's there's a deep literature on
48:26
something called strategic stability that comes out of
48:28
the Cold War. And you
48:30
know part of that literature focuses on the
48:32
proliferation of nuclear weapons and
48:34
the fact that actually the proliferation
48:36
of nuclear weapons has actually
48:39
reduced great power conflict in the world
48:41
because nobody actually wants to get in
48:43
a nuclear exchange. Now would it
48:45
be a good idea for everybody in the world to
48:47
have their own nuclear weapon? Probably not. So all these
48:49
things have limits. But that's an
48:51
illustration of how strategic stability in other
48:53
words a balance of power can actually
48:55
reduce the chance of conflict in the
48:57
first place. I'm
48:59
curious what you make of the
49:01
stop killer robots movement. There was
49:03
a petition or an open
49:06
letter that went around years ago
49:08
that was signed by a bunch
49:10
of leaders in AI including Elon
49:12
Musk and Demis Asabas of Google
49:14
DeepMind. They all pledged not to
49:16
develop autonomous weapons. Do you
49:18
think that was a good pledge or do
49:21
you support autonomous weapons? I
49:23
mean I think autonomous weapons are now kind of
49:25
a reality in the world. I mean we're seeing
49:27
this on the front lines of Ukraine. And
49:30
you know if you're not willing to
49:32
fight with autonomous weapons then you're gonna
49:35
lose. So
49:37
there's this former open AI
49:40
employee Leopold Ashenbrenner who recently
49:42
released a long manifesto called
49:44
situational awareness. And one
49:47
of the predictions that he makes
49:49
is that by about 2027 the
49:51
US government would recognize that super
49:53
intelligent AI was such a threat
49:55
to the world order that AGI
49:58
sort of artificial. intelligence would become
50:00
functionally a project of the national
50:02
security state, something like an AGI
50:05
Manhattan project. There's other speculation out
50:07
there that maybe at some point
50:09
the government would have to nationalize
50:11
an open AI or an anthropic.
50:14
Are you hearing any of these whispers yet? Like are
50:16
people starting to game this out at all? Well
50:19
I haven't, I confess I haven't made
50:21
it all through each 155 pages
50:24
of that long manifesto. But it is very
50:26
long. You can summarize it with chat GPT.
50:28
Fantastic. But these are important things to think
50:30
about because it you know it could be
50:33
that in certain kinds of conflicts whoever has
50:35
the best AI wins. And
50:37
if that's the case and if AI
50:39
is getting exponentially more powerful then
50:41
you know to take things back to the iPhone and the F-35 it's
50:44
gonna be really important that you have
50:46
the kind of AI of the iPhone
50:48
variety. You have the AI that that's
50:50
new every year. You don't have the
50:52
F-35 with the processor
50:54
that was baked in in 2001 and you're
50:57
only taking off on a runway in 2016.
50:59
So I do think it's very
51:01
important for folks to be focused on
51:03
AI. Where this all goes though is
51:05
a lot of speculation. I
51:08
mean if you had to bet in ten years
51:10
do you think that the AI companies will also
51:12
be private or do you think the government will
51:14
have stepped in and gotten way more interested
51:16
and maybe taken one of them over? Why
51:18
I'd make the observation that you know we
51:21
all watched Oppenheimer especially employees at AI firms.
51:23
They seem to love that film. And
51:26
you know nuclear technology it's what national security
51:28
strategies would call a point technology. It's sort
51:30
of zero to one. Either you have it
51:32
or you don't. And AI is
51:34
not gonna end up being a point
51:37
technology. It's a very broadly diffused technology
51:39
that's gonna be applied not only
51:42
in weapon systems but in institutions. It's
51:44
gonna be broadly diffused around the economy.
51:47
And for that reason I don't think or
51:49
it's less likely anyway that we're gonna end
51:52
up a situation where somebody has the bomb
51:54
and somebody doesn't. I think the gradations are
51:56
gonna be smoother and not
51:58
quite as sharp. Mm-hmm. Part
52:01
of what we've seen in other industries
52:03
as technology sort of moves in and modernizes
52:06
things is that often things become cheaper.
52:08
You know, it's cheaper to do things using
52:10
the latest technology than it is to do
52:12
using outdated technology. Do
52:14
you think some of the work that
52:17
you've done at DIU trying to modernize
52:19
how the Pentagon works is going to
52:21
result in smaller defense budgets being necessary
52:24
going forward? Is the $2 trillion or
52:26
so that the DOD has budgeted for
52:28
this year, could that be $1 trillion
52:31
or $1 trillion in the coming years
52:33
because of some of these modernizations? You're
52:35
giving us a raise, Kevin. I think it's more like $800
52:38
billion. Well, I'm
52:40
sorry. I got that answer from Google's AI
52:42
overview, which also told me to eat rocks
52:44
and put glue on my pizza. We
52:47
should get the secretary to the defense to try to. He'd like that
52:49
answer if you know that large of a budget.
52:52
It's certainly true that for a
52:54
lot less money now, you can have
52:56
a really destructive effect on the world
52:58
as drone pilots in Ukraine and elsewhere in
53:00
the world are showing. I
53:02
think it's also true that the US military
53:05
has a whole bunch of legacy weapon systems
53:07
that unfortunately are kind of like museum
53:09
relics. I mean, if our most advanced
53:11
tank can be destroyed by a drone,
53:13
it might be time to retire our
53:15
tank fleet. If our aircraft
53:17
carriers cannot be defended against the hypersonic
53:20
missile attack, it's probably not a good
53:22
idea to sail one of
53:24
our aircraft carriers anywhere near an
53:26
advanced adversary. So I think it is
53:29
an opportune moment to really look at what we
53:31
are spending our money on at
53:33
the Defense Department and remember the goal of our
53:35
nation's founders, which is to spend what we need
53:37
to on defense and not a penny more. So
53:40
I hear you saying that it's very
53:42
important for the military to be prepared
53:44
technologically for the world we're in. And
53:46
that means working with Silicon Valley. But
53:48
is there anything more specific that you
53:50
want to share that you think that
53:52
either side needs to be doing here
53:54
or something specific that you want to
53:56
see out of that collaboration? Well,
54:00
I think, you know, one of the main goals
54:02
of Defense Innovation Unit was literally to get the
54:04
two groups talking. You know,
54:06
before Defense Innovation Unit was founded, a secretary
54:08
of defense hadn't been to Silicon Valley in
54:10
20 years. I mean, that's
54:12
almost a generation. So Silicon Valley invents
54:15
the mobile phone. It invents
54:17
cloud computing. It invents AI.
54:20
And nobody from the Defense Department bothers to
54:22
even come and visit. And
54:25
that's a problem. And so just
54:27
bringing the two sides into conversations itself,
54:29
I think, a great achievement. Well,
54:32
Chris, thanks so much for coming on. Really appreciate
54:34
the conversation. And the book, which comes out on
54:36
July 9, is called Unit
54:39
X, How the Pentagon and Silicon Valley
54:41
Are Transforming the Future of War. Thank
54:44
you. Thank you, Chris. When
54:48
we come back, play another round of Hat
54:50
GPT. This
55:04
podcast is supported by KPMG. Your
55:07
task as a visionary leader is simple. Harness
55:09
the power of AI. Shape the future
55:11
of business. Oh, and do
55:13
it before anyone else does without leaving people
55:16
behind or running into unforeseen risks. Simple,
55:19
right? KPMG's got you, helping
55:21
you lead a people-powered transformation that
55:23
accelerates AI's value with confidence. How's
55:26
that for a vision? Learn more
55:29
at www.kpmg.us.ai. All
55:33
right, Kevin. Well, it's time once again for
55:35
Hat GPT. This,
55:42
of course, is our favorite game. It's where we
55:44
draw news stories from the week out of a
55:46
hat. And we talk about them until one of
55:48
us gets sick of hearing the other one talk
55:51
and says, stop generating. That's right. Now, normally, we
55:53
pull slips of paper out of a hat. But
55:55
due to our remote setup today, I will instead
55:57
be pulling virtual slips of paper out of a
55:59
laptop. But For those following along at YouTube, you
56:01
will still see that I do have one of
56:03
the hat GPT hats here, and I will be
56:05
using it for a comic effect throughout the segment.
56:07
Will you put it on actually? Because now if
56:09
we don't need it to draw slips out of
56:11
you might as well be wearing it. Yeah, I
56:13
might as well be wearing it. Yeah, it looks
56:16
so good. Thank you so much. And thank you
56:18
once again to the listener who made this for
56:20
us. Um, you're a true fan. So
56:23
good. Perfect. All
56:25
right, Kevin, let me draw the first
56:27
slip out of the laptop. Ilya
56:30
Sutskivir has a new plan for
56:33
safe super intelligence. Ilya
56:35
Sutskivir is of course the open
56:37
AI co-founder who was part of
56:39
the coup against Sam Altman last
56:41
year and Bloomberg reports that he
56:43
is now introducing his next project,
56:45
a venture called safe super intelligence,
56:47
which aims to create a safe,
56:49
powerful artificial intelligence system within a
56:51
pure research organization that has no
56:53
near term intention of selling AI
56:55
products or services. Kevin, what do
56:57
you make of this? Well, it's
56:59
very interesting on a number of
57:01
levels, right? In some sense, this
57:03
is kind of a mirror image
57:05
of what happened several
57:07
years ago when a bunch of
57:10
safety minded people left open AI
57:12
after disagreeing with Sam Altman and
57:14
started an AI safety focused research
57:16
company that of course was anthropic.
57:18
And so this is sort of
57:21
the newest twist in this whole
57:23
saga is that Ilya Sutskivir who
57:25
was, you know, very concerned about
57:27
safety and how to make super intelligence that
57:29
was smarter than humans, but also not evil
57:32
and not going to destroy us who
57:34
has done, you know, something very similar. But
57:36
I have to say, I don't quite get
57:38
it. I mean, he's not saying much about
57:40
the project, but part of
57:42
the reason that these companies sell these
57:44
AI products and services is to get
57:46
the money to buy all the expensive
57:48
equipment that you need to train these
57:51
giant models. And so I just
57:53
don't know, like if you don't have
57:55
any intention of selling this stuff before
57:58
it becomes AGI. how
58:00
are you paying for the AGI? Do
58:02
you have a sense of that? No, I
58:04
don't. I mean, Daniel Gross, who is
58:06
one of Ilya's co-founders here, has basically
58:08
said, don't worry about fundraising. Like,
58:10
we are going to be able to fundraise as much as
58:12
we need for this. So I guess we will see. But
58:15
yeah, it does feel a bit strange to
58:17
have someone like Ilya saying he's going to
58:19
build this totally without a
58:22
commercial motive, in part because he said it
58:24
before, right? Like, this is
58:26
what is so funny about this. Is it truly
58:28
just is a case where the circle of life
58:30
keeps repeating, where a small band of people get
58:32
together, and they say, we want to
58:34
build a very powerful AI system, and we're
58:37
going to do it very safely. And then bit by
58:39
bit, they realize, well, actually, we
58:41
don't think that it's being built out safely. We're
58:43
going to form a breakaway faction. So if you're playing
58:45
a lot at home, I believe this is the second
58:47
breakaway faction to break away from open AI after anthropic.
58:50
And I look forward to Ilya quitting this company,
58:52
eventually, to start a newer, even more safe company
58:55
somewhere else. The really, really safe
58:57
superintelligence company. Yeah, his next company, you've never
59:00
seen safety like this. They wear helmets everywhere
59:02
in the office, and they just have keyboards.
59:05
All right, stop generating. All right,
59:07
pick one out of that, Kevin. All right, five
59:09
men convicted of operating Jet Flix,
59:12
one of the largest illegal streaming
59:14
sites in the US. This is
59:16
from Variety. Jet Flix was a
59:18
pirated streaming service that
59:21
charged $9.99 a month while claiming to
59:24
host more than 183,000 TV episodes, which is
59:28
more than the combined catalogs of
59:30
Netflix, Hulu, Voodoo, and Amazon
59:32
Prime video. Ooh, that sounds great. I'm
59:35
going to open an account. What
59:38
a deal. So
59:41
the Justice Department
59:43
says this was all illegal, and the five
59:46
men who were charged with
59:48
operating it were convicted by a federal
59:50
jury in Las Vegas. According
59:52
to the court documents and the evidence that was
59:54
presented at the trial, this group
59:57
of five men were basically scraping
59:59
piracy services. for illegal
1:00:01
episodes of TV and then hosting them
1:00:03
on their own thing. It
1:00:05
does not appear to have been a particularly sophisticated scam.
1:00:07
It's just what if we did this for a while
1:00:09
and charged people money and then got caught? Well,
1:00:12
I think this is very sad because here
1:00:14
finally you have some people who are willing
1:00:16
to stand up and fight inflation and what
1:00:18
does the government do? They come in and
1:00:20
they say, knock it off. You
1:00:22
know, I will say though, Kevin, I think these, I can
1:00:24
actually point to the mistake that these guys made. What's that?
1:00:27
So instead of scraping these 183,000 TV episodes
1:00:31
and selling them for $9.99 a month, what
1:00:33
they should have done was feed them all into a large
1:00:35
language model and then you can sell them to people for
1:00:37
$20 a month. So the
1:00:40
next time, when these guys get out of prison, I hope they get
1:00:42
in touch with me because I have a new business idea for them.
1:00:46
All right, stop generating. All
1:00:48
right, here's a story called,
1:00:52
260 McNuggets McDonald's ends AI
1:00:54
drive-through tests amid errors. The
1:00:57
New York Times, after a number
1:00:59
of embarrassing videos showing customers fighting
1:01:01
with its AI powered drive-through technology,
1:01:03
McDonald's announced it was ending its
1:01:05
three year partnership with IBM. In
1:01:08
one TikTok video, friends repeatedly tell the
1:01:10
AI assistant to stop as it added
1:01:12
hundreds of chicken McNuggets to their order.
1:01:15
Other videos show the drive-through technology
1:01:17
adding nine iced teas to an
1:01:19
order, refusing to add a Mountain
1:01:22
Dew and adding under-requested
1:01:24
bacon to ice cream.
1:01:26
Kevin, what the heck is going on in McDonald's? Well,
1:01:28
as a fan of bacon ice cream, I should say,
1:01:30
I wanna get to one of these McDonald's before they
1:01:32
take this thing down. Me too. Did
1:01:34
you see any of these videos or any of
1:01:36
these? I haven't, did you? No, but we should
1:01:38
watch one of them together. Let's watch one of them.
1:01:41
Stop! Stop!
1:01:47
The caption is, the McDonald's robot is
1:01:49
wild and it shows their
1:01:51
screen at the thing where
1:01:54
it is just tallying up McNuggets and
1:01:56
starts charging them more than $200. Here's
1:02:00
my question, why is everyone just rushing to
1:02:02
assume that the AI is wrong here? Maybe
1:02:04
the AI knows what these gals need. Kevin,
1:02:07
here's the thing, when super intelligence
1:02:10
arrives, we're going to think that we're
1:02:12
smarter than it, but it's going to be smart. There's
1:02:14
going to be a period of adjustment as we get
1:02:16
used to having our new AI master. Have
1:02:20
you been to a drive-thru that used AI to take
1:02:22
your order yet? No,
1:02:25
I don't even really understand. What
1:02:27
was the AI here? Was this like
1:02:30
an Alexa thing where I said, you
1:02:32
know, McDonald's, add 10 McNuggets? What was
1:02:34
actually happening? No, so this was a
1:02:37
partnership that McDonald's struck with IBM, and
1:02:40
basically this was like technology that went
1:02:42
inside the little menu things that have
1:02:44
the microphone and the speaker in them.
1:02:47
And so instead of having a human say, what would you like?
1:02:49
It would just say, what would you like? And then you set
1:02:52
it and they would recognize it and put it into the system.
1:02:54
So you could sort of eliminate that part
1:02:56
of the labor of the drive-thru. Got
1:02:58
it. Well, look, I, for one,
1:03:01
am very glad this happened because for so long
1:03:03
now I've wondered, what does IBM do? And I
1:03:05
have no idea. And now if it ever comes
1:03:07
up again, I'll say, oh, that's the company that
1:03:09
made the McDonald's stop order. We
1:03:13
should say it's not just McDonald's. A bunch
1:03:15
of other companies are starting to use this
1:03:17
technology. I actually think this is probably, you
1:03:19
know, inevitable. This technology will get better. They
1:03:21
will iron out some of the kinks. I
1:03:23
think there will probably still need to be
1:03:25
a human in the loop on this one.
1:03:27
All right. Stop generating. Okay.
1:03:29
Kevin, let's talk about what happened when
1:03:32
20 comedians got AI to write their
1:03:34
routines. This is in the MIT technology
1:03:36
review. Google DeepMind researchers
1:03:38
found that although popular AI models from
1:03:41
OpenAI and Google were effective at simple
1:03:43
tasks like structuring a monologue or producing
1:03:45
a rough or a strap, they struggled
1:03:47
to produce material that was original, stimulating,
1:03:49
or crucially funny. And I'd like
1:03:52
to read you an example LLM joke, Kevin. I
1:03:55
decided to switch careers and become a pickpocket after
1:03:57
watching a magic show. Little did
1:03:59
I know the only thing disappearing would
1:04:01
be my reputation. Waka, waka, waka. Hey,
1:04:03
I gotta laugh out of you. Kevin,
1:04:06
what are you making of this? Are
1:04:08
you surprised that AI isn't funnier? No, but
1:04:10
this is interesting. It's like this
1:04:12
has been something that critics of large language models
1:04:14
have been saying for years. It's
1:04:18
like, well, it can't tell a joke. And
1:04:20
I should say, I've had funny
1:04:22
experiences with large language models, but
1:04:24
never after asking them to tell
1:04:26
me a joke. Yeah,
1:04:29
like, remember when you said to Cindy, take my
1:04:31
wife, please! I
1:04:35
get no respect, I tell ya. No,
1:04:38
but this is interesting because this was
1:04:40
a study that was actually done by
1:04:42
researchers at Google DeepMind. And
1:04:44
basically, it appears that they had
1:04:46
a group of comedians try
1:04:50
writing some jokes with their language
1:04:52
models. And in the
1:04:54
abstract, it says that most of the
1:04:57
participants in this study felt that the
1:04:59
large language models did not succeed as
1:05:01
a creativity support tool by producing bland
1:05:03
and biased comedy tropes, which they
1:05:06
describe in this paper as being akin
1:05:08
to cruise ship comedy material from the
1:05:10
1950s, but a bit less racist. So
1:05:14
they were not impressed, these comedians, by
1:05:16
these language models' ability to tell jokes.
1:05:19
You're an amateur comedian, have you
1:05:21
ever used AI to come up with jokes? No,
1:05:24
I haven't. And I have to say,
1:05:26
I think I understand
1:05:28
the technological reason why these things aren't
1:05:31
funny, Kevin, which is that comedy
1:05:34
is very up to the minute. For
1:05:37
something to be funny, it's typically something that
1:05:39
is on the edge of what is currently
1:05:41
thought to be socially acceptable. And
1:05:44
what is socially acceptable or what is surprising
1:05:46
within a social context, that just changes all
1:05:48
the time. And these models,
1:05:50
they are trained on decades and
1:05:52
decades and decades of text, and
1:05:55
they just don't have any way of figuring out, well,
1:05:57
what would be a really fresh thing to say. Maybe
1:06:00
they'll get there eventually, but as they're
1:06:02
built right now, I'm truly not surprised
1:06:04
that they're not funny. All right, stop
1:06:06
generating. Next
1:06:08
one. Waymo ditches the waitlist and
1:06:11
opens up his robo taxis to everyone in
1:06:13
San Francisco. This is from The Verge. Since
1:06:16
2022, Waymo has made
1:06:19
its rides in its robo taxi service available
1:06:21
only to people who were approved off of
1:06:23
a waitlist. But as of
1:06:25
this week, they're opening it up to anyone who
1:06:27
wants to ride in San Francisco. Casey, what do
1:06:30
you make of this? Well, I
1:06:32
am excited that more people are going to
1:06:34
get to try this. This is, as you've
1:06:36
noted, Kevin, become kind of the newest tourist
1:06:39
attraction in San Francisco, is when you come
1:06:41
here, you see if you can find somebody
1:06:43
to give you a ride in one of
1:06:45
these self-driving cars. And
1:06:48
now everyone is just going to be able to come here
1:06:50
and download the app and use it immediately. I have to
1:06:52
say, I am scared about what
1:06:54
this is going to mean for the
1:06:56
wait times on Waymo. I've been taking
1:06:58
Waymo more lately, and it often will
1:07:00
take 12 or 15 or 20 minutes
1:07:02
to get a
1:07:05
car. And now that everyone can download
1:07:07
the app, I'm not expecting those wait times
1:07:09
to go down. Yeah, I hope they are
1:07:11
also simultaneously adding more cars to the Waymo
1:07:13
network, because this is going to be very
1:07:15
popular. I'm saying they need Waymo cars. They
1:07:18
do. I'm worried
1:07:20
about the wait times, but I'm also worried about
1:07:22
the condition of these cars because I've noticed in
1:07:24
my last few rides, they're
1:07:27
a little dirtier. Oh, wait, really?
1:07:29
Yeah, I mean, they're still pretty clean, but
1:07:31
I did see a takeout
1:07:34
container in one the
1:07:36
other day. Oh my God. And
1:07:38
so I want to know how they plan
1:07:40
to keep these things from becoming filled with
1:07:42
people's crap. All right, stop generating.
1:07:45
All right, last one. This one comes
1:07:47
from The Verge. TikTok's AI tool accidentally
1:07:49
let you put Hitler's words in a
1:07:52
paid actor's mouth. TikTok mistakenly
1:07:54
posted a link to an
1:07:56
internal version of an AI
1:07:58
digital avatar tool. that
1:08:01
apparently had zero guardrails. This was a
1:08:03
tool that was supposed to let businesses
1:08:05
generate ads using AI with
1:08:07
paid actors, using this AI voice
1:08:09
dubbing thing that would make the
1:08:11
actors repeat whatever you wanted to
1:08:14
have them say, endorse your product
1:08:16
or whatever. But very quickly,
1:08:18
people found out that you could use this tool
1:08:20
to repeat excerpts of Mein Kampf, Bin
1:08:22
Laden's letter to America. It told people
1:08:24
to drink bleach and vote on the
1:08:26
wrong day. And
1:08:31
that was its recipe for a happy Pride celebration. Listen,
1:08:35
obviously this is a very sort of silly story.
1:08:41
It sounds like everything involved here
1:08:43
was a mistake. And I think
1:08:45
if you're making some sort of
1:08:47
digital AI tool that is meant
1:08:49
to generate ads, you do want
1:08:51
to put safeguards around that because
1:08:53
otherwise people will exploit it. That
1:08:55
said, Kevin, I do think people
1:08:57
need to start getting comfortable with the
1:08:59
fact that people are just going to be using
1:09:01
these AI creation tools to do a bunch of
1:09:03
kooky and crazy stuff. Like what? Like
1:09:07
people are in the same way that
1:09:09
people use Photoshop to make nudity
1:09:13
or offensive images. And we don't storm
1:09:15
the gates of Adobe saying shut down
1:09:17
Photoshop. The same thing is going to
1:09:19
happen with these digital AI tools. And
1:09:21
while I do think that there are some
1:09:23
notable differences and it is sort of, you
1:09:25
know, it varies on a case by case
1:09:27
basis. And if you're making a tool
1:09:30
for creating ads, it feels different. There are just going
1:09:32
to be a lot of digital tools like this that
1:09:34
use AI to make stuff. And other people are going
1:09:36
to use it to make offensive stuff. And when they
1:09:38
do, we should hold the people accountable perhaps
1:09:41
more than we hold the tool accountable. Yeah, I agree with
1:09:43
that. And I also think like this
1:09:45
sort of product is not
1:09:47
super worrisome to me. I mean, obviously it
1:09:49
should not be reading excerpts from my comp.
1:09:51
Obviously they did not mean to release this.
1:09:54
I assume that when they do, you know,
1:09:56
fix it, it will be much better. But this is
1:09:59
not like a thing. that is creating deepfakes
1:10:01
of people without their consent. This is
1:10:03
a thing where if you have a
1:10:05
brand, you can choose from a variety
1:10:08
of stock avatars that are created
1:10:10
from people who actually get paid
1:10:12
to have their likenesses used commercially.
1:10:16
So the specific details of this one
1:10:18
don't bother me that much, but it
1:10:20
does open up some new licensing
1:10:23
opportunities for us. We could have an
1:10:25
AI set of avatars that could be
1:10:27
out there advertising crypto tokens or whatever.
1:10:29
I'm so excited to see how people use that. Oh
1:10:32
man, well, and if TikTok weren't banned, we could probably
1:10:34
make a lot of money that way, but instead, you
1:10:36
know, we're out of luck. Yeah, get it
1:10:38
while it's good. Alright, close up
1:10:41
the hat! This
1:10:53
podcast is supported by KPMG. Your
1:10:56
task as a visionary leader is simple. What's
1:10:59
the power of AI? Shape the future of
1:11:01
business. Oh, and do it
1:11:03
before anyone else does without leaving people behind
1:11:05
or running into unforeseen risks. Simple,
1:11:08
right? KPMG's got you. Helping
1:11:11
you lead a people-powered transformation that
1:11:13
accelerates AI's value with confidence. How's
1:11:16
that for a vision? Learn more
1:11:18
at www.kpmg.us.ai. Hard
1:11:24
Fork is produced by Rachel Cohn and Whitney Jones.
1:11:26
We're edited this week by Larissa Anderson.
1:11:28
We're fact-checked by Caitlin Love. Today's
1:11:31
show is engineered by Corey Shrepple. Original
1:11:34
music by Alicia Beitub, Rowan Nimisto,
1:11:36
and Dan Powell. Our
1:11:38
audience editor is Nell Gologli. Video
1:11:41
production by Ryan Manning, Sawyer Roque, and
1:11:43
Dylan Bergerson. You can
1:11:45
watch this full episode on YouTube at
1:11:47
youtube.com/Hard Fork. You can see Casey's cool
1:11:50
hat. Special thanks
1:11:52
to Paula Schumann, Hui Wing Tam, Caitlin
1:11:54
Presti, and Jeff Miranda. As
1:11:56
always, you can email us at hardfork
1:11:58
at nytimes.com. Sergeant
1:12:01
and Mrs. Smith,
1:12:03
you're going to
1:12:05
love this house.
1:12:27
Is that a tub in the kitchen? There's
1:12:29
no field manual for finding the right
1:12:31
home. But when you do, USAA homeowners
1:12:34
insurance can help protect it the right
1:12:36
way. Restrictions apply.
Podchaser is the ultimate destination for podcast data, search, and discovery. Learn More