Episode Transcript
Transcripts are displayed as originally observed. Some content, including advertisements may have changed.
Use Ctrl + F to search
0:07
I want you to think back to almost exactly
0:09
a year ago. All of
0:11
a sudden, one name, one product
0:14
really, was everywhere. It's
0:16
called ChatGPT, which stands for
0:18
Generative Pre-trained Transformer, and it's
0:20
fully powered through artificial intelligence.
0:22
This project from the Open
0:24
AI Research Lab can write
0:26
essays and carry on convincing
0:28
written conversation. It took Netflix
0:30
more than three years to reach
0:32
one million users, but it took
0:34
ChatGPT just five days. Techies
0:37
everywhere short-circuiting with excitement. ChatGPT
0:39
is a disruptor and a
0:41
game changer for business communication.
0:44
Computers have achieved a sort of creativity.
0:46
Conversational AI is a tool to help us learn
0:49
faster, apply it in the right way, and there
0:51
are billions to be made. We
0:55
were talking about ChatGPT. You
0:57
were talking about ChatGPT. Your
0:59
not-very-online relatives were talking about
1:02
ChatGPT. And
1:04
according to Karen Howe, that moment was
1:06
an inflection point. When
1:08
ChatGPT came out, it was the first consumer-facing
1:11
demonstration of really powerful AI
1:14
capabilities that suddenly made everyone
1:16
in the public, all
1:18
policymakers, everyone's mom,
1:21
grandma, like you know, uncle, suddenly
1:23
come online to the idea that
1:25
this technology is a really big
1:27
deal and it's going to have
1:29
massive cascading effects all around society.
1:32
The reason I wanted to talk to
1:34
Karen is that she knows more about
1:36
AI than probably any other reporter in
1:38
the country. She's covered the
1:40
industry for years and is now writing a book
1:43
about it. She says
1:45
the release of ChatGPT changed everything
1:47
for the company that created it,
1:49
OpenAI, and for every
1:51
other AI company that hadn't yet brought something
1:53
like it to the public. And
1:56
in that moment, it was also
1:58
a huge glaring block. red
2:01
flag to every other company within
2:04
the world that has the resources
2:06
to develop this technology to hurry
2:08
up quick and start doing something
2:10
similar. It also
2:12
helps explain why it was international
2:14
news when OpenAI's board ousted CEO
2:17
Sam Altman and he then
2:19
clawed his way back. Because
2:21
this company, more than any other,
2:23
has upended what Silicon Valley means
2:25
when it says AI. And
2:28
so in the moment when OpenAI
2:30
released this change of tea, it basically changed the entire game,
2:33
it's changed suddenly all
2:35
of the, basically the
2:37
orientation of the entire tech industry to
2:41
consolidate around this singular idea of
2:43
let's try to use this technology to build
2:45
this kind of cap-on-like thing or to build
2:47
so-called large language models. And
2:50
that's really not happened
2:52
before. We haven't seen in the
2:54
entire trajectory of AI development, we
2:56
haven't seen a moment that so
2:59
quickly made everyone
3:01
start doing the same exact
3:03
thing. But,
3:11
and this is important, there was a catch.
3:14
A contradiction baked into the founding
3:16
of OpenAI. The organization
3:18
began as a nonprofit with a
3:20
mission to help humanity. Only
3:22
a small part of the company was supposed to
3:25
make a profit and a cap-on at that. The
3:28
board, however, maintained the priorities
3:30
of the nonprofit. Karen
3:33
says a fight between the two sides was almost
3:35
inevitable. What's interesting
3:38
about this whole fiasco
3:40
that happened is I'm
3:42
of the belief that the board kind
3:44
of did its job, like that's exactly what
3:47
it was set up to do. The whole
3:50
reason for this kind of weird mechanism
3:52
at the time was essentially like an
3:54
elaborate way to try to self-regulate. and
4:00
was in fact, if they
4:02
believed at some point that
4:05
Sam Allman or whoever the CEO was, was
4:09
leading the company astray from the original
4:11
mission to create technology that was beneficial
4:14
to everyone, that the board would
4:16
then have the right to fire them. And
4:19
it's sort of interesting. I think there's
4:21
a lot of rightful criticism about the
4:23
way that the board went about it,
4:25
but I do think that the fact that
4:27
they did it itself should
4:30
not be criticized because that's what they
4:32
signed up for. That's what everyone signed
4:34
up for. And in fact, Sam Allman
4:36
himself was a key author of this legal
4:38
structure that gave the board the power to do
4:40
this to him. But
4:46
now that the dust has settled, Sam Allman
4:48
clearly won the site. He's
4:50
back in control. The original board is
4:52
gone. And this idea
4:54
of self-regulation in the AI business
4:56
seems almost painfully naive, considering the
4:59
amount of money at stake. Today
5:02
on the show, the open AI fiasco
5:04
is a morality play for the industry.
5:07
One with massive stakes for all of
5:09
us as AI development races ahead. I'm
5:12
Lizzie O'Leary, and you're listening to What Next TBD,
5:15
a show about technology, power, and how the
5:17
future will be determined. Stick around.
5:32
Karen told you about the structure of open AI,
5:35
a nonprofit with a for-profit wing.
5:38
That wing, the money-making side, brought
5:40
chat GPT to market and
5:42
attracted $13 billion in investment
5:45
from Microsoft. But open
5:47
AI has more than just a split structure.
5:50
There's also an ideological split among the
5:52
people who work there and in the
5:54
field at large. You can
5:56
call them the techno optimists and the
5:58
AI doomers. so
7:32
extreme like the police around
7:34
these two different camps has
7:36
become so stream is
10:00
the most aggressive company in terms
10:02
of commercialization. Now, the company
10:04
is launching product after product, not
10:07
just Chat GPT, but the image
10:09
maker Dolly, and a whole host
10:11
of services for AI startups. They
10:13
recently hosted a conference for developers in
10:16
which they touted a platform to create
10:18
custom AI chatbots. Think of
10:20
it like an app store for your own Chat
10:22
GPT. We're thrilled to introduce
10:26
GPTs. GPTs
10:28
are tailored versions of Chat GPT for
10:31
a specific purpose. You
10:33
can build a GPT, a customized
10:36
version of Chat GPT, for almost anything,
10:38
with instructions, expanded knowledge,
10:41
and actions. And
10:43
then you can publish it for others to use. It's
10:45
sort of the most extreme
10:47
Silicon Valley of Silicon Valley
10:50
ideas. It's
10:53
sort of like the recreation of the Apple
10:56
App Store, or the turning of a product
10:58
into a platform. All of these really deeply-seated
11:01
Silicon Valley methods.
11:03
Before the hearing of Altman
11:06
happened, OpenAI
11:08
was starting to become widely criticized. What
11:10
is your nonprofit board even doing? What
11:13
is it there for? Because the
11:15
for-profit arm was essentially just taking
11:17
the lead and just seemed to
11:19
look exactly like any other startup,
11:22
any other company. And
11:24
now that Altman has been
11:26
reinstated, arguably,
11:29
he is going to make a huge effort
11:32
to try and make sure he never has
11:34
this kind of vulnerability again, where he can
11:36
get fired by the board again. And
11:38
so there were a lot of, clearly, the
11:40
board and the
11:43
OpenAI executives and Altman during
11:45
this window of time had
11:47
huge negotiations, very
11:50
tense negotiations, about how
11:52
to actually create some kind of
11:55
viable exit plan that everyone agreed with.
11:57
And ultimately, it was a huge effort.
12:00
I think Altman ended up giving quite a
12:02
lot of concessions in that he's no longer
12:04
on the board So that is one sign
12:06
that he doesn't have like full Power
12:10
in this scenario in like the new
12:12
kind of iteration of open AI But
12:15
certainly he is going to
12:17
try like every other possible way in
12:20
addition to the board Who
12:22
continue pushing ahead his own vision
12:24
and making sure that he can
12:26
continue to pursue what he wants
12:28
whether that's? Commercialization or not. Do
12:31
you think it's fair to say that? In
12:35
this reorganization He now
12:37
has more power. I would
12:39
guess Yes But
12:42
a lot of it is going to
12:44
come down to how much
12:46
these board members Turn out
12:48
to ally with him the new ones the
12:51
new ones And I think
12:53
all men would not have agreed to
12:55
these board members and for himself
12:57
to step off the board If
13:00
he didn't think that they were Big
13:03
potential allies, but we won't
13:05
really know I mean up until
13:08
the board fired Altman I also
13:10
thought that the previous board was
13:12
pretty allied with all men So
13:15
it was pretty surprising to me that they would
13:17
end up doing an action as drastic
13:19
as this Although again, it was in their job
13:21
description But
13:24
you know, we won't really know whether
13:26
these new board members might Take
13:31
a similar approach until they actually act on
13:33
it But for for the time
13:35
being the best guess is that Altman wouldn't have
13:37
allowed them on the board unless he personally felt
13:40
very confident That he would be able to turn
13:42
them into allies if they aren't already one
13:45
of the reasons I'm asking that question is because
13:49
We don't still know kind
13:51
of the fine-grained details of
13:53
why the original board Fired
13:56
him they put out a statement that seemed
13:58
sort of dramatic in the moment and
14:01
we have these competing camps that
14:03
you've illustrated, but
14:06
there's always the possibility
14:08
that more details
14:10
could drip out. And so
14:12
it makes me wonder whether we
14:15
should expect those little breadcrumbs to come in
14:17
the next few months, or if,
14:19
nope, door is closed, let's just move on
14:22
with this new version of the company. I
14:25
think if we see details coming out, it
14:27
will have to be from the previous board,
14:31
or the current board, actually, that made
14:33
the decisions, because ultimately, the previous
14:36
board never communicated to employees
14:40
why they made the decisions that they did.
14:42
The employees themselves are also in the dark.
14:44
And of course, I don't think Altman is
14:46
going to himself reveal, or
14:48
who knows, maybe he would reveal because it
14:50
might be strategically savvy, but
14:53
we're not gonna find out for employees.
14:55
There's only a small, a tiny
14:57
handful of people that truly know
14:59
the reasons, and it will
15:01
be up to them to speak. And I think
15:05
right now, my sense from
15:07
speaking with sources is that
15:09
the board and
15:12
executives, and Altman in particular,
15:14
are really trying to project
15:16
a sense of stability because
15:19
it's not just ultimately about his image
15:23
and what the public thinks of him, but also,
15:25
OpenAir obviously has Microsoft as a huge investor.
15:28
Right, they put $13 billion into this. Exactly,
15:32
so there's just a lot of incentives
15:34
right now for everyone to project stability
15:36
and unity, and to give this impression
15:38
that everything is fine, and we're gonna
15:40
continue chugging along, and that was a weird blip.
15:43
But, you know, ultimately, we
15:47
could see more details dripping out as
15:50
things maybe settle. When
15:56
we come back, what's the AI doomers are
15:59
really afraid of? legal
18:00
structure was an
18:03
experiment in the best
18:05
possible way that you could self-regulate, and
18:08
it still fell apart. I mean,
18:10
that's a huge damning sign that
18:12
this whole thing... I mean, we've been talking
18:14
about this for so long that self-regulation
18:17
doesn't work, but this is
18:19
yet another example, a cherry
18:21
on the top of mounting evidence.
18:25
And I hope that policymakers
18:27
are aware of this. I hope that
18:29
consumers and the general public become more
18:31
aware that, oh, wait, this is actually
18:33
just, in a sense, another
18:37
manifestation of Silicon Valley. And
18:40
AI is not just this organic technology
18:43
that emerges, but it's actually created by
18:45
people with particular beliefs and particular
18:47
agendas, and that
18:50
can help inform them in making
18:53
better decisions of consumers
18:55
about how much they want to incorporate
18:58
this technology into their lives, into their
19:00
work, into their... For doctors,
19:02
into their medical practices or for teachers,
19:04
into their classrooms. And I
19:06
think that is the bigger lesson that I
19:08
hope people can all take away. I
19:11
wonder if you could articulate for someone who
19:13
is outside of Silicon Valley, who doesn't think
19:15
about this on a constant basis, what
19:19
those fears are about? What
19:22
does the existential risk look
19:25
like? What does the day-to-day risk
19:27
look like? I
19:29
think the existential risk fears
19:32
are based on this
19:34
hypothetical that if
19:37
we do believe that these
19:39
systems are intelligent, which
19:41
is already entering
19:44
kind of a quagmire because... Right, like
19:46
what is intelligence? Yeah,
19:48
there is no scientific consensus around what intelligence
19:50
is. But if we believe that
19:52
these systems are intelligent, digital
19:55
intelligences are just...
19:57
They're faster at learning, faster at
19:59
learning. at combining knowledge,
20:01
knowledge in quotes, than
20:04
humans. So humans, when
20:06
we have knowledge and we learn,
20:09
it's sort of a very inefficient process. It
20:11
takes us many, many years to get like
20:14
a proper education. And then when we talk
20:16
to each other, we don't really combine knowledge
20:18
very effectively because people will have disagreements or
20:20
different interpretations about things. But with
20:22
a digital, quote unquote,
20:24
intelligence, they would be able to
20:27
learn from data, like within a
20:29
few months, and then just transfer
20:32
the data that they trained on and
20:34
the things that they learned instantly. And
20:38
so under this premise or under this belief
20:40
that this is true, then
20:42
you could see why it could be
20:46
very scary because you would be
20:48
able to quickly create a
20:50
super intelligence that is smarter than humans
20:52
on many different tasks and could not
20:55
just be smarter, but outsmart
20:58
humans and start to manipulate
21:00
people and create sort of
21:02
a life of its own where its objective
21:04
function is to perpetuate its own existence. And
21:06
if that means humans get in the way
21:08
of that, then destroy humans. I mean, again,
21:12
this is like an extremely hypothetical
21:14
scenario and is based on a
21:16
lot of different assumptions about what
21:18
intelligence is, whether or not it's
21:21
successfully being created within these digital technologies
21:24
and also whether or not
21:26
these so-called digital intelligences could
21:29
even act in the physical
21:31
world because again, they're digital and part
21:33
of the reason why humans can
21:35
do things in the world and have
21:37
potentially dangerous consequences in the world is
21:39
because we have physical bodies and we
21:41
walk around in three dimensional space. The
21:44
other camp kind of of
21:47
people that are concerned about risks of
21:49
AI is often they often call
21:52
it like short term risks or risks that are
21:54
in the here and the now. And
21:56
these people are concerned about the things that
21:58
AI technologies that we are already have
22:01
and what we've already seen, like real
22:03
examples that we've already seen of ways
22:05
that they break down and can cause
22:07
harm to people. So for this
22:10
group, they're like, we shouldn't even be talking about
22:12
what intelligence is and whether or not
22:15
we're recreating it. We should just talk
22:17
about like the literal things that we
22:19
have and observe the fact that self-driving
22:21
cars have killed pedestrians and
22:24
observe that facial recognition systems
22:26
have been involved in the wrongful arrest of
22:29
black men in the U.S. and
22:32
observe that there have been hiring
22:34
algorithms developed that don't hire women
22:36
because they learn over time that
22:39
that is sort of the best
22:41
way to maximize for
22:43
certain types of traits that
22:45
don't necessarily represent what we
22:47
need to maximize as a society. And
22:51
so that kind of camp is
22:53
much more, I mean,
22:55
there's a lot of tension between these camps because you
22:57
could see that for the
22:59
existential risk people, they would
23:02
argue that like you need
23:05
to project further into the future than what
23:07
we see now and that
23:09
when you project into the future, of course, it's
23:11
going to be hypothetical. Whereas the
23:13
people that are focused on the short term
23:15
risks are like, actually, if we focused
23:18
on solving the problems with AI
23:20
today, that would naturally get us
23:23
to better AI in the
23:25
future and to overlook the
23:27
things that are happening today
23:29
and just project hypothetically. You're
23:31
not actually going to get
23:33
anywhere meaningful because you're not
23:35
actually recognizing how
23:37
AI literally interfaces with society.
23:40
A lot of what you're talking about reminds
23:44
me of the way
23:46
products, any tech product, is
23:50
a product of the people who
23:52
built it. Their smarts, their
23:55
biases, what
23:57
they're deeply versed in, what they're not.
24:00
These are phenomenally complex and
24:02
powerful systems. When
24:05
I think about the new board
24:07
of OpenAI, it
24:09
is all men. Does
24:11
that give you pause that there
24:14
is so much
24:16
power in the hands of such
24:18
a small group of people? It
24:21
gives me pause whether or not they're all men
24:23
or not. Say more. It's
24:26
interesting that they are
24:28
all men because I would say if
24:30
there had been one woman on the board,
24:32
I mean, even before there were two women on the board, if
24:35
there had been one woman on the board, if there
24:37
had been one non-white person on the board, that
24:40
it almost would have given a false sense
24:42
of security and a false sense of representation
24:45
because it would belie the
24:47
fact that there are still only
24:49
three of them. And
24:53
that fundamentally is the bigger problem.
24:57
And for me, I do
24:59
think that obviously it is
25:02
problematic that these three people do not
25:04
seem to represent the diversity
25:07
of society, but could they have ever?
25:11
And I think that is the bigger question.
25:13
And I hope that actually the fact that there is
25:15
no representation on the board
25:17
kind of accelerates that bigger
25:20
discussion. You've read
25:22
this other story that I have been thinking
25:24
about. It was back in October
25:26
and the headline was, we don't actually know
25:28
if AI is taking over everything. And
25:31
I wanted to talk about that a little bit because I recently
25:34
was in San Francisco, you get off
25:36
the plane, it's like every other poster
25:38
or billboard is for an AI
25:40
company, literally everything. And
25:43
if you step back from that, it seems harder
25:46
to define just how
25:49
transformative AI is or isn't
25:52
because as you wrote about, very
25:54
little of it is transparent about
25:57
how it does what it does.
26:01
How can we know how
26:03
much AI is impacting our lives already?
26:06
We don't. This
26:09
is a huge problem in
26:12
that it is very much driven by
26:14
the fact that AI development is now
26:18
happening primarily in secrecy
26:20
within these companies that have huge
26:23
incentives to continue keeping that secrecy
26:25
because of competitive advantages and also
26:27
because OpenAI says, oh, we can't
26:29
tell you because
26:31
it would be dangerous to society
26:33
for people to know this. I
26:36
mean, what we do know is that, OK,
26:38
I'll nuance it a bit in that
26:41
we do know that AI
26:43
is having a huge impact on society.
26:46
What we don't know is whose
26:49
AI is the
26:51
most pervasive and where they are
26:53
pervasive, which is a very
26:55
important detail because it
26:58
matters if OpenAI's
27:01
AI is everywhere or
27:03
if another kind of AI is everywhere because
27:06
that helps us scrutinize, understand,
27:08
and hold accountable ultimately
27:10
the developers of the technology that are
27:13
impacting us. If we don't know what
27:15
AI our doctor is using, what AI
27:17
our teacher is using, what AI our
27:19
lawyer is using, we can't
27:22
know how much to trust
27:25
a particular result. We
27:27
don't know who to contest
27:29
if something is wrong. And
27:33
regulators, fundamentally, then can't
27:35
create rules that make
27:37
sense and apply
27:39
to specific cases of how this
27:41
technology is used, the entire supply chain
27:44
of how it's used. So
27:46
that is, I think, the more fundamental
27:48
problem is the who. We just don't
27:50
know the who. And your
27:53
life could be entirely run by
27:55
OpenAI's algorithms. And you would not
27:57
know that. And that is the
27:59
problem. And yet
28:01
I think if you are a person
28:04
whose life does not directly touch the
28:06
tech industry or Silicon Valley, you might
28:08
be thinking, this is
28:10
really complicated. It seems
28:13
too hard to get my head around right
28:15
now. What
28:19
should you be thinking about
28:22
so that you don't find yourself
28:24
in a situation where AI
28:28
is deeply enmeshed in your life before you
28:30
realized it? I think
28:32
asking lots of questions and keeping
28:34
your eyes open to the possibility
28:37
that AI could be used in
28:39
many, many ways that you
28:41
don't necessarily know about. If you are
28:43
a parent, you can definitely
28:47
ask questions with the school, with
28:49
your kids' teachers about, are you using AI
28:51
in the classroom? How are you using it
28:53
in the classroom? Who are you using? And
28:57
if you are, I mean, anyone who
29:00
lives in a city with a local government
29:02
could be asking their local government
29:05
officials, like, how are you thinking
29:07
about AI? How are you
29:09
potentially going to acquire some of these
29:11
tools? In which agencies, in which
29:14
processes are you going to incorporate these
29:16
tools? And what kind of accountability mechanisms
29:18
are you going to put in place?
29:21
I think that kind of
29:23
just speaking up
29:25
and observing and asking
29:27
those questions to map out
29:30
for yourself how your
29:33
life might be
29:35
affected by these technologies is
29:37
a really critical first step to then figuring out
29:39
whether or not you want them to be a part
29:42
of them, a part of your life. And
29:45
I think also just for
29:47
consumers, I mean, a lot of people now use
29:49
chatGBC, a lot of people use BARD, a lot
29:51
of people use Anthropics
29:53
Quad. I
29:56
also think just like thinking more about how
29:59
you actually want to incorporate. of this technology and like
30:02
any other products that we use,
30:05
how we want to vote with our money, which
30:07
companies do we want to give that money to,
30:09
and to put
30:11
pressure with that money
30:14
when a company does not actually
30:17
align with our expectations of
30:19
being a good actor in the world. Karen
30:28
Howe, thank you so much for
30:30
your insight and your time. Thank
30:32
you so much, Lizzie. Karen
30:35
Howe is a contributing writer for The
30:37
Atlantic and is writing a book for
30:39
Penguin Press about the AI industry. And
30:42
that is it for the show today. What Next
30:45
TBD is produced by Evan Campbell and Anna Phillips.
30:48
Our show is edited by Jonathan Fisher. Alicia
30:50
Montgomery is vice president of audio for
30:52
Slate. TBD is part of
30:54
the larger What Next family, and we're
30:57
also part of Future Tense, a partnership
30:59
of Slate, Arizona State University, and New
31:01
America. And if you're a fan
31:03
of the show, I have a request for you. Join
31:06
Slate Plus. Just head
31:08
on over to slate.com/what next plus to
31:10
sign up. It also makes a great
31:12
holiday gift for the newshound in your
31:14
life. We will be
31:16
back on Sunday with another episode. I'm Lizzie
31:19
O'Leary. Thanks for listening.
Podchaser is the ultimate destination for podcast data, search, and discovery. Learn More