Latest / The Smart Cookies podcast / Why AI Won't Replace Workers, But Will Crash The Economy. With Cory Doctorow

Transcript

Why AI Won't Replace Workers, But Will Crash The Economy. With Cory Doctorow · The Smart Cookies podcast

0:00
And when the AI bubble pops, we're going to see
0:02
a massive economic vacuum, right? 35 % of the
0:05
stock market vaporized. We will likely see governments
0:08
defaulting to austerity, which has the well -known
0:12
effect of driving support for fascism, right?
0:15
When you lose the things that matter to you,
0:17
your doctor, your job, your school. you become
0:21
a sucker for people who show up and say the reason
0:24
you've lost the things that matter is that undeserving
0:26
others have taken it and you should embark on
0:29
racist pogroms, which we see taking place everywhere.
0:32
We see austerity. So we'll see more austerity.
0:35
We'll see a breakdown in our systems. I'm absolutely
0:41
delighted to welcome back journalist, blogger,
0:43
author, and friend of the show, Corey Doctorow.
0:45
A pleasure to be talking to you again, sir. Thank
0:48
you very much. Pleasure to be back on. You recently
0:50
appeared on a lot of really high profile programs
0:54
and the comments under these videos are pretty
0:56
unanimous. Most people are saying this is the
0:59
voice we need right now. Additionally, I have
1:01
to confess, much as it's hard for a cold hearted
1:05
Eastern European like me that I've never received
1:07
so much private correspondence urging me to have
1:10
someone back on the pod. than I did after our
1:13
last conversation. So before we talk about your
1:15
new book, I wanted to ask you, why do you think
1:18
your analysis of the internet and AI and of big
1:21
tech in general is resonating with so many folks
1:24
right now? Well, I've been at this for 25 years.
1:29
So I work for the Electronic Frontier Foundation.
1:31
It's a digital rights group based in America,
1:34
but I was their inaugural European director and
1:36
I worked on their behalf in more than 30 countries.
1:39
And having done this in 30 countries for 25 years,
1:43
I like to think I have some context that rather
1:47
than treating all of these issues as though they
1:50
occur in a vacuum, like we were just hit by a
1:52
meteor called AI or, you know, a meteor hit Silicon
1:57
Valley 10 years ago and created in shitification.
2:00
I can tell you where these things came from.
2:03
I can tell you how we can get rid of them. I
2:05
don't have to resort to mystical explanations
2:09
like Mark Zuckerberg's the smartest person that
2:11
ever lived and, you know, succeeded where Rasputin
2:14
and Mesmer and MKUltra and pickup artists failed
2:17
and he built a mind control ray. And that's why
2:20
everyone's acting so weird because Facebook's
2:22
got a mind control ray. I can ground this in
2:25
material and economic phenomena. I can tell you
2:28
why we have bubbles. I can tell you why my fellow
2:32
progressives are wrong when they say endless
2:34
growth is the ideology of a tumor. The reason
2:37
they want endless growth in Silicon Valley has
2:39
nothing to do with ideology. It's about material
2:41
factors. And I think that when you can ground
2:45
the phenomena around you that are making you
2:48
angry and sad and scared in real material things,
2:51
that because they're real material, suggest ways
2:57
that we can resolve them. That even if you're
3:00
still angry and sad and upset at the end of that,
3:04
at least you know what to do, right? At least
3:07
you're not scared in the sense that you're a
3:09
prisoner of history. Maybe you're scared because
3:12
now you understand the forces arrayed against
3:14
you and what needs to be done to deal with them.
3:18
I mean, Elon Musk also has a lot of context,
3:21
but not even his followers are claiming that
3:24
he's the voice of reason. Well, I think Elon
3:28
Musk makes people follow, makes people excited
3:33
about him because he deals exclusively in a kind
3:37
of atavism and in kind of libidinal emotions.
3:40
And while there's something libidinal and something
3:43
atavistic about words like and shitification
3:45
and the minor license to vulgarity and moments
3:48
where everything seems to be going wrong. I like
3:51
to think that I really. respect the people who
3:55
take the time to listen to me enough to assume
3:57
that they have the native intelligence to follow
4:01
a nuanced argument that doesn't assume that Jews
4:04
are importing people of color to outbreed white
4:08
people and end Western civilization or any of
4:12
the other absurd things that Elon Musk believes,
4:14
including that USAID doesn't do anything important.
4:19
All right, you mentioned incentivization. a couple
4:21
of times now. We talked about it last time and
4:23
the term really took off. It's everywhere. I
4:26
can only imagine how hard some Google execs jaws
4:29
must be clenching when they hear this word being
4:32
applied to one of their products. Did you ever,
4:34
I'm just wondering, did you ever get like a sassy
4:37
email or anything from any of these people? I
4:40
mean, I sometimes hear from them. Mostly what's
4:43
happened is I've had a couple of high profile
4:45
speaking gigs canceled because one of the sponsors.
4:48
And no one's ever told me who the sponsor was.
4:50
But, you know, you look at the sponsor roles
4:52
and it's all the usual suspects. So I think that's
4:56
the that's how that is being felt. And, you know,
5:00
I guess it's always a mystery to me why someone
5:04
would invite me to speak at a conference that's
5:06
being sponsored by firms that I have a long track
5:08
record of saying. intemperate but true things
5:13
about that make people angry at those companies.
5:16
You know, more power to them for inviting me.
5:19
I'm grateful for it. But I think that there is
5:21
this very foreseeable outcome, I think, in many
5:24
cases that those sponsors are going to object
5:27
eventually. You know, I think 20 years ago, 25
5:32
years ago, the sector was diverse enough. There
5:36
were enough players in it. that there was always
5:38
a kind of balance of power within it. When I
5:41
look at institutions like the W3C, the World
5:43
Wide Web Consortium that designed standards for
5:45
the web, and I look back in those days when there
5:47
were hundreds of members of almost the same size.
5:51
And so the worst impulses of the members kind
5:54
of died in their tracks because there would be
5:56
other members who had better impulses. And I
5:58
look now where when there's a group within the
6:01
W3C of very large firms that want to do something
6:04
that's obviously bad for the web, they win. It's
6:07
not because the W3C is weak or foolish or evil.
6:10
It's because when your budget depends on a dozen
6:14
of your highest paying members and they all decide
6:16
that they want one thing, it's very hard to resist
6:19
them. You've since written a new book, The Reverse
6:24
Centaur's Guide to Life After AI, How to Think
6:27
About Artificial Intelligence Before It's Too
6:29
Late. I know you've had to explain the concept
6:31
of a reverse centaur a gazillion times already,
6:35
but for anyone who might be, unfamiliar still.
6:38
Would you be so kind as to break it down again?
6:40
Yeah. So it won't surprise you to know that this
6:43
is not the first time that labor and automation
6:46
have come into conflict with one another. And
6:48
so there's a pretty good literature explaining
6:50
how different kinds of automation scenarios have
6:54
played out in labor markets and for the people
6:56
who rely on the outputs of labor, you know, the
6:59
things that we get from workers who use machines.
7:02
And one of the... I guess iron laws or truisms
7:06
or cornerstones of that theory is that when workers
7:09
adopt automation, they typically do so in order
7:12
to improve quality. But when bosses drive automation
7:15
adoption, it's typically in service to increasing
7:18
throughput. Because, you know, if you are paying
7:21
a monthly subscription or if you've acquired
7:23
an asset that's now depreciating, you want to
7:25
maximize your use of it, right? That's how you
7:27
maximize your profit, right? You've got an expense
7:30
now. You want to... make the most of it. You
7:33
want to, as they say, sweat the asset. And this
7:36
is particularly true where you have situations
7:39
where you have monopolies or cartels or duopolies
7:42
that control the market because they can make
7:44
worse quality products and still sell them because
7:48
they're the only game in town. And so this is
7:51
what you see when the dollar general that's replaced
7:54
the grocery store in your town makes everything
7:56
smaller and more expensive. What are you going
8:00
to do? take the bus 10 miles to the last remaining
8:03
grocery store. So in those conditions, you would
8:08
expect things to get worse. And so this is where
8:11
centaurs and reverse centaurs come in. In automation
8:14
theory, a centaur is someone who is assisted
8:17
by a machine. The analogy here is the human head
8:21
on the horse's body. The body being the body
8:23
of a horse, it's tireless, it's strong, it's
8:26
fast, but it doesn't have discernment or judgment.
8:30
That comes from the human head. So the human
8:32
head is directing this strong, fast, enduring
8:36
system, machine, body. So when you're riding
8:40
a bicycle, you're a centaur, right? Even look
8:42
like one. When you're using a spell checker,
8:44
you're a centaur too, you know. There are lots
8:47
of ways in which we can become centaurs. A reverse
8:50
centaur, the corollary, that's the reverse, right?
8:53
That is the horse's head on the human body. That's
8:55
the machine driving the human. That's the machine
8:58
making decisions about the human. And of course,
9:00
the machine is still stronger than you, faster
9:03
than you. It has more endurance than you. There's
9:05
just some part of the task that it can't do.
9:08
And so you are its peripheral and you are its
9:12
help me. And because you are slower and weaker
9:15
and get tired more easily, it is going to work
9:19
you at the limit of your endurance and at the
9:21
limit of your capability. And that means that
9:23
you're not just going to get used by the machine.
9:25
You're going to get used up. by the machine.
9:27
And we see this in our kind of memes and tropes
9:31
about automation and labor. You know, people
9:33
who've seen I Love Lucy will know this classic
9:35
scene where Lucy and Ethel are trying to put
9:37
chocolates in the chocolate box and the conveyor
9:39
belt's going so fast that the chocolates are
9:42
going everywhere. Or that classic scene in modern
9:45
times where Charlie Chaplin gets sucked in the
9:47
machine and ground up in the gears. You know,
9:49
this is that. logical consequence of capital
9:53
-driven automation, an inferior product that
9:56
makes the workers who produce its lives hell.
9:59
And it's why the most automated warehouses we've
10:02
ever built, the Amazon warehouses, also have
10:04
the highest rate of injury of any warehouse in
10:06
the sector. It's not a coincidence. It's the
10:08
consequence, right? If you're spending seven
10:10
or eight figures on automation, you're not going
10:11
to run the machine slower than you need to. You
10:13
will run them at the limit of the worker's capacity.
10:17
And of course, if you're working at your limit
10:18
for long stretches, you will eventually make
10:20
a mistake. And when you make that mistake in
10:22
a warehouse, sometimes you get impaled on a forklift.
10:27
The last part of the title of the book is even
10:29
scarier than what you were just talking about.
10:32
I'm not going to lie. What do you mean by before
10:35
it's too late? Well, look, AI is a bubble, which
10:39
is not to say that AI technology isn't useful
10:42
if labor is driving automation. adoption here
10:45
if labor was in charge of ai i'm sure we'd see
10:47
lots of people finding useful things for it in
10:49
fact when you meet programmers who say oh i write
10:52
code with ai and it's the best code i've ever
10:54
written and then you meet other programmers who
10:57
are writing code with ai and they say i can't
10:58
believe how bad this code is you know i work
11:01
in avionics and i don't think anyone should ever
11:03
get on an airplane because no one has ever produced
11:05
tech debt at the scale that we're producing it
11:08
now with these ai tools the way to understand
11:10
that contradiction is not that the first group
11:12
is lying or the second group is lying. It's that
11:15
the first group are centaurs and the second group
11:16
are reverse centaurs. So there are useful things
11:19
we can do with AI, but the bubble where they've
11:22
raised now over $1 .4 trillion on less than $50
11:26
billion a year in revenue, that bubble demands
11:29
that we fire. as many workers as possible and
11:32
replace them with defective chatbots. And even
11:34
if we do that, we're still going to see a rupture.
11:37
That bubble is going to burst. 35 % of the US
11:40
stock market, the seven companies. are AI companies,
11:43
and they trade around the same $100 billion IOU
11:46
really fast, pretending it's in all their bank
11:48
accounts at once. Every generation of AI technology
11:51
loses more money than the previous generation.
11:53
Every new AI customer loses money for the business.
11:57
Every time an AI customer touches the AI, the
11:59
business loses even more money. And so here we
12:03
are replacing critical functions in our society
12:05
with technology that will eventually just be
12:08
switched off. We are wiping out the kind of hard
12:15
won expertise, what's called sometimes process
12:17
knowledge, right? All the things that workers
12:20
know that can't really be written down, that
12:22
tacit knowledge that's held between workers within
12:25
a firm, across different divisions and across
12:27
different firms in a sector and up and down a
12:29
supply chain. All of that stuff's being wiped
12:31
out because we're firing those workers and they're
12:33
becoming discouraged or they're retraining or
12:35
they're retiring. And when the AI bubble pops.
12:38
we're going to see a massive economic vacuum,
12:41
right? 35 % of the stock market vaporized. We
12:43
will likely see governments defaulting to austerity,
12:47
which has the well -known effect of driving support
12:50
for fascism, right? When you lose the things
12:52
that matter to you, your doctor, your job, your
12:55
school. You become a sucker for people who show
12:59
up and say the reason you've lost the things
13:01
that matter is that undeserving others have taken
13:03
it and you should embark on racist pogroms, which
13:06
we see taking place everywhere. We see austerity.
13:09
So we'll see more austerity. We'll see a breakdown
13:12
in our systems. And it's going to be really hard.
13:18
And the sooner we burst the bubble, the better.
13:21
And if we wait too long. Some of these AI companies
13:25
are going to unload on regular investors. Right
13:28
now, most of that AI money is big, dumb money.
13:31
It's institutionals. It's Gulf states. It's billionaires.
13:36
But, you know, those people are really itching.
13:40
to get you to relieve them of their soon -to
13:43
-be worthless stock. We just saw this with Elon
13:45
Musk and SpaceX, where he did a sweetheart deal
13:48
with the Nasdaq 100 to debut SpaceX in the index
13:52
of top Nasdaq shares. That's never done. Normally,
13:56
if a company debuts in the stock market, it's
13:59
ineligible for inclusion in an index until it's
14:01
had several quarters of sustained profitability,
14:04
because it might just be like a... pump and dump.
14:07
It might be a swindle. And those indexes are
14:10
held by everyday savers. That's what most people
14:12
put their retirement savings into is an index
14:14
fund. And if you manage an index fund, you are
14:17
legally obliged to buy stock from every company
14:20
in the index. So if SpaceX is in the top 100,
14:24
you have to go buy SpaceX. What Musk did when
14:27
he floated SpaceX is he didn't just get this
14:30
exemption to the restriction on debuting in the
14:33
index. He also restricted the shares that the
14:36
company put in up on the market. So the company
14:39
put almost no shares up for sale. All of the
14:42
shares that were sold to fund managers who were
14:45
required to buy them at any price with your retirement
14:50
savings, all those shares came from insiders.
14:54
who were not obliged to follow the normal legal
14:57
requirement of a lockout where insiders aren't
15:00
allowed to sell for the first 90 days or 180
15:02
days, again, in case it's a pump and dump. And
15:05
so, you know, we see these, these companies already
15:08
starting to unload everything on ordinary savers
15:11
to create basically bag holders. And again, that
15:15
like there. It's not just the injustice, although
15:19
this is very important, or the social problem
15:21
of people having their retirement savings wiped
15:24
out. It's also what it does to the trusted institutions
15:27
and to our social cohesion. It paves the way
15:30
for authoritarian movements when people lose
15:32
everything. And so that's the real risk that
15:36
we face right now, is the collapse of our...
15:39
important firms and institutions because we'll
15:41
replace skilled workers with AI and then the
15:43
AI will be shut off when the companies run out
15:45
of money. And that will make people furious because
15:50
they will enter a period of austerity right after
15:54
having all of their savings stolen by stock swindlers
15:57
who are also insiders. And that's going to be
15:59
just catastrophic. What do you mean AI is just
16:03
going to be switched off? Because after the dot
16:06
-com... bubble burst the internet wasn't just
16:09
like switched off and a lot of people are saying
16:10
yeah yeah no that's very true so the dot coms
16:14
had very different unit economics right so that
16:16
a lot of them were dumb right for sure but as
16:19
a sector the web had great unit economics every
16:22
new web customer made more money for web companies
16:25
every time you use the web the web made more
16:28
money and every generation of the web made more
16:31
money than the previous generation of the web
16:32
that's not how ai works as i've said ai is the
16:36
opposite of that and so while there will be and
16:39
i this is actually quite important there's going
16:41
to be a ton of gpus kicking around there's going
16:43
to be like data centers looking for buyers at
16:45
fire sale prices there's going to be skilled
16:47
workers who know how to do things with ai who
16:49
don't have to answer to their stupid bosses anymore
16:52
and there's going to be these open source models
16:54
that have barely been optimized and that whenever
16:56
anyone tries to optimize them they get all kinds
16:58
of of new efficiencies out of them that stuff
17:01
will be around but like No one's going to be
17:04
keeping OpenAI around in its current form. OpenAI
17:07
is a business that sells $100 bills for $1 each,
17:10
right? And even if you wipe out all of its capital
17:15
expenditure overhang and start over again through
17:18
a bankruptcy, if you continue to give away $100
17:22
bills for $1 each, you will quickly end up exactly
17:25
where you started. So it might be that you can...
17:30
replace a worker who costs you $100 ,000 a year
17:34
with a chatbot that costs you $1 ,000 a year
17:36
and not really care that the outputs are half
17:39
as good as the worker produced. But if the chatbot
17:42
costs $120 ,000 a year and the outputs are half
17:46
as good, that's not the same situation at all.
17:49
And that's what we would be in if we were going
17:51
to continue to consume AI resources at the rate
17:55
we're consuming them now, but actually had to
17:57
pay for them instead of having them defrayed
17:59
by Gulf state sovereign wealth funds and Silicon
18:03
Valley billionaires. Yeah, I'm no economist,
18:06
but selling $100 bill for $1 doesn't sound like
18:10
it. Oh, and it gets worse than that. When the
18:13
companies were trying to clean up their balance
18:14
sheets because they were all talking about their
18:16
IPOs last month, they were like, what if we start
18:18
charging $5 for these $100 bills? And all the
18:23
CEOs who bought AI for their businesses and said,
18:26
everyone has to use AI. And if I catch you not
18:28
using AI, I'm going to fire you. They turned
18:30
around. They said, these $100 bills are not worth
18:33
$5. I'm sorry. We just don't get enough use out
18:36
of them. uh to pay five dollars for them we are
18:39
going to institute you know company -wide bans
18:42
on using ai limits on using ai you're gonna have
18:45
to ask your manager for more tokens if you use
18:47
up the tokens that you've been given right so
18:49
like there it's not just that they're selling
18:52
hundred dollar bills for a dollar it's that no
18:55
one will pay more than a dollar for their hundred
18:57
dollar bills it's really bad I want to talk about
19:01
the ideology behind this AI push a little bit.
19:04
Now, I totally believe that the engineers working
19:06
on these LLMs have probably read too much science
19:09
fiction and are working earnestly towards creating
19:13
some sort of benevolent superintelligence or
19:16
whatever. However, when it comes to their bosses,
19:18
the CEOs who are steering this thing, do you
19:21
reckon that in their minds this whole project
19:24
is basically... given what we're talking about
19:27
right now, just like every capitalist's wet dream?
19:30
So a campaign to simply eliminate human labor?
19:35
I think that there's two things going on here.
19:38
I think for sure some of these guys definitely
19:40
would like to live in a world without people.
19:42
I think there's a kind of solipsism that goes
19:45
with being a billionaire. You can't really inflict
19:48
the kind of pain at scale that you have to in
19:51
order to make a billion dollars or hundreds of
19:53
billions of dollars. If you really believe other
19:56
people feel the way you feel are as real as you
19:59
are, you know, and that's true of this class
20:02
when they're face to face and in individual interactions.
20:05
So, you know, the Epstein class just couldn't
20:08
have believed that the children who are on that
20:09
island are as real as their own children. But
20:12
it's also true, especially when they deal with
20:13
people as a statistical abstraction. You know,
20:16
Jeff Bezos and his army of drivers and warehouse
20:19
workers who don't even get pee breaks. Right.
20:21
Like he would never. stipulate to those indignities.
20:25
He gets a catheter fitted when he goes up on
20:28
his spaceship. He doesn't just pee in a bottle.
20:32
So even under the most extreme circumstances,
20:35
there's lots of toilets on his yacht. The idea
20:38
that you don't get a pee break is something for
20:40
other people and not for him. And so you get
20:43
Elon Musk. calling everyone who disagrees with
20:46
them an NPC and so on. So I think there's a lot
20:49
of these people who just don't think other people
20:50
are as real as they are, or at least, you know,
20:53
everyday people. And certainly, like, if you
20:55
run a social media business, I think it's probably
20:57
really bad. Like Mark Zuckerberg. deals with
20:59
people entirely as statistical abstractions.
21:01
You know, he's just like twisting dials in the
21:05
back end of Facebook to see if he can evince
21:07
human behaviors at scale. Right. So this is really
21:11
a disindividuated way of confronting people.
21:13
But then you have on top of that, you have what
21:17
Milton Keynes called the beauty contest. So Milton
21:20
Keynes said that making money by investing. is
21:24
not a matter of picking the most beautiful person
21:26
in the beauty contest the company that's going
21:28
to make the most money making money as an investor
21:31
is about picking the person the other judges
21:34
will think is most beautiful because you are
21:37
trying to pick the stock that other people are
21:40
going to want to buy irrespective of whether
21:42
you think that stock is good so you know there's
21:45
people who invest in you know the supplements
21:48
that andrew tate and uh you know those other
21:53
chud podcasters joe rogan and so on that oh yeah
21:57
you know there's people who invest in those in
22:00
those um supplements who don't believe that the
22:03
supplements work, right? They know that it's
22:05
just a way of getting extremely expensive urine,
22:07
but they also know that there's like an infinite
22:11
army of vulnerable, desperate young man who will
22:13
pay everything they have for the supplements.
22:15
You don't have to think the supplements will
22:16
work to invest in the supplements. You don't
22:18
have to think that, you know, Gwyneth Paltrow
22:20
has a jade egg that will make your vagina better
22:23
to think that. Women will buy the jade egg and
22:26
risk their lives by putting it in their bodies.
22:29
Right. And so I think by the same token, there's
22:32
a lot of people who think, you know, bosses are
22:34
really easy sells for a sales pitch about how
22:38
you can replace workers with software because
22:43
like workers are a pain in the ass. They demand
22:45
moral consideration. And even more so, they provoke
22:49
these kind of. ego -shattering confrontations
22:52
where you have an idea, you want to do a thing,
22:56
and then you have to go out and talk to a worker
22:58
who knows how that thing works. And they tell
23:01
you that your idea is stupid or illegal or impossible
23:05
or it's going to kill people. And eventually
23:07
you start to have this suspicion that while you're
23:10
telling yourself you're in the driver's seat,
23:12
the fact is that if you didn't show up for work,
23:15
everything would be fine. And if they didn't
23:17
show up for work, everything would stop. And
23:19
maybe you're in the backseat with a Fisher -Price
23:21
steering wheel. And so AI is like really the
23:24
potential of like wiring the steering wheel directly
23:26
into the drivetrain, right? You have an idea
23:29
and it just gets shit out by the AI. And so,
23:32
you know, you don't have to believe that AI can
23:34
do that to believe that bosses will buy it if
23:36
you promise them that. And, you know, you don't
23:39
have to be able, you don't have to be fast enough
23:40
to outrun the bear. You just have to be fast
23:42
enough to outrun everyone else. So long as you
23:44
get out of the AI position before. everyone realizes
23:48
that the ai can be useful for workers who want
23:50
to augment their work but not particularly useful
23:53
for replacing workers and you get out ahead then
23:57
it doesn't matter that it it was a bad investment
24:00
it doesn't matter that the company failed you
24:03
got out okay right there's plenty of people who
24:06
made a lot of money on nfts nfts were are and
24:10
always will be stupid and a bunch of people made
24:12
a lot of money from nfts I mean, it's bizarre
24:17
how they're not even trying to come up with a
24:20
sensible justification for all this, are they?
24:22
It's just a half -baked version of a trust -me
24:24
-bro. They're saying we have to throw everything
24:27
into their company so they get to AGI as fast
24:30
as possible, and then out of the goodness of
24:32
their hearts, they will reap the benefits on
24:34
the rest of humanity. Now, call me Mr. Cynical,
24:37
but when in the history of the planet did it
24:39
ever happen that the guy at the top, and usually
24:42
it was a guy... said, you know what, now that
24:44
I robbed all of you blind, I think I'm going
24:46
to divide the pie fairly. Let everyone have a
24:49
nice piece. Or maybe I'm, you know, just not
24:52
as well read on my history as I thought I was.
24:54
Well, you know, there's a connection between
24:56
this and some of the more esoteric forms of billionaire
24:58
mysticism. So, you know, the Sam Bankman Freed
25:02
and a bunch of his cronies practiced a thing
25:04
called effective altruism. I've heard of it,
25:06
yeah. Yeah, it has its roots in some pretty benign
25:09
stuff, like trying to empirically determine whether
25:11
you save more lives investing in like anti -malaria
25:14
research or mosquito nets. Right. But it quickly
25:18
became this kind of hyper utilitarian self -parodying
25:22
venture where you. earn to give so you take the
25:27
most destructive harmful job you can that pays
25:30
the most to earn as much money as possible no
25:33
matter how much pain you inflict on people around
25:35
you in the name of earning so much money that
25:38
you can give it to causes that will make people's
25:40
lives better But of course, the way to improve
25:43
the most lives is not to help people who are
25:46
alive today, but rather to invest all your money
25:49
in improving the lives of 10 to the 53 artificial
25:53
humans who will come into existence on Venus
25:56
in 10 ,000 years. Because if you make even the
25:59
smallest improvement to the lives of 10 to the
26:01
53 people, it will swamp all the benefit you
26:04
could possibly confer. on a mere 8 billion people
26:08
who are alive today, right? This is like, this
26:10
is a very weird idea, right? And like, it is,
26:14
you know, as they say, I think, again, this is,
26:18
I can't remember if this is Keynes or Galbraith,
26:19
but, you know, it's the endless pursuit of man's
26:24
oldest ideological project to find a superior
26:26
moral justification for selfishness, right? So
26:29
much of this, you know, kind of cashes out there.
26:35
I think that that it's worse than the idea that
26:40
they're just hand waving and saying, trust me,
26:42
bro. And like using the shitty trope from science
26:45
fiction novels where computers get complicated
26:47
enough and then they wake up, which is like what
26:49
science fiction writers do because we're lazy
26:51
and we just want to get straight to the futuristic
26:53
parable. We don't want to come up with a bunch
26:56
of ways in which computers can become self -aware.
26:58
We just say like, oh, you add enough transistors
27:00
and now it's a brain. Right. It's worse than
27:04
that. When you read accounts of the NIH researchers
27:07
who pleaded with the Doge boys to not cancel
27:12
their long -running cancer research, cancel their
27:16
long -running cancer research. Sorry, it's hard
27:18
to say. What the Doge kids said was, we don't
27:23
need cancer research because general AI is around
27:26
the corner. And when general AI arrives, it will
27:29
cure cancer. So it's worse than that. It's not
27:33
just that they're making false promises. It's
27:36
that they're discontinuing the actual promising
27:38
things. If you want a more immediate version
27:40
of this, every time an AI bro turns on a gas
27:44
turbine and is asked about the climate impact,
27:46
they say, don't worry, when General AI comes
27:48
along, it'll solve the climate emergency. And
27:51
our job now is to outrun. the second long of
27:55
thermodynamics, right? Before we melt the ice
27:58
caps, we have to make the general AI that tells
28:02
us what to do once the ice caps melt, right?
28:05
It's an incredible gamble they are making with
28:10
the planet and their basis for asserting that
28:13
a general AI is around the corners is deeply...
28:17
foolish. They're basically saying that if you
28:20
teach enough words to the word guessing program,
28:22
it will wake up, which is like saying that if
28:25
you breed horses to run fast enough, one of them
28:27
will give birth to a locomotive, right? Like
28:30
it doesn't matter how fast the horse runs. It's
28:32
not a train and it doesn't matter how well the
28:35
program guesses words. It's not conscious. Consciousness
28:38
is not just guessing words. Okay. Since we're
28:43
here, let's talk about the state of of this technology
28:45
today. Because for somebody like me, he usually
28:47
uses it as some sort of like an expanded Google
28:50
search. I honestly have no idea if this thing
28:52
is like the philosopher's stone or just a really
28:55
confident parrot. And the discourse at large
28:58
is really not helping me. For example, when I
29:00
opened my ex yesterday, God help me, my feed
29:02
was suddenly full of folks saying that AI has
29:04
cracked some old mathematical problem. And apparently
29:07
it's true. It's not bullshit. On the other hand,
29:10
immediately after that, I watched a video on
29:12
the work of Jennifer Dowd. She's a scientist
29:15
who I want to have on the podcast someday. She's
29:17
a pioneer in gene editing, and she said that
29:19
chatbots are basically useless when it comes
29:22
to innovation, and that she thinks it's going
29:23
to stay that way for a while. So which of the
29:26
two is it? I'm confused. You know, one of the
29:30
things that AI, and this isn't the first generation
29:32
of AI we've ever had. AI was coined at a conference
29:35
in the 1950s, and we've had many different kinds
29:37
of AI emerge. And, you know, while they're all
29:41
radically different and while AI has been a marketing
29:44
term for each of them, what they all share in
29:46
common is that they were all a bid to replicate
29:50
and possibly even replace some function that
29:53
we think of as innately human. And what each
29:56
generation of AI has done is showed us that while
29:59
there is a uniquely human way to do certain things,
30:03
that there are other ways to do it, too. Right.
30:05
The way that humans play chess is not the way
30:08
that AI plays chess. Uh, and, uh, while they
30:12
may look the same from the outside, they're doing
30:15
something different and you can achieve the state
30:18
of a chess game with AI, like a successful chess
30:21
game, a victorious chess game in a way that's
30:24
different from the way people do it, which is
30:26
interesting, super interesting, right? I mean,
30:29
for a long time, you look at the history of,
30:31
of, uh, um, human flight. And for a long time,
30:35
we were kind of becalmed in those doldrums of
30:38
trying to mimic birds. We just built ornithopters,
30:40
which are basically planes that flap their wings.
30:43
And they never worked. And we had to invent a
30:47
totally different way of flying to actually leave
30:52
the ground. There's a famous quote about thinking
30:55
machines where a computer scientist was asked,
30:59
can machines think? And they answered, can submarines
31:02
swim? Submarines don't swim, right? But they
31:07
move through the water. They do some of the things
31:08
we do with thinking. You know, I don't know if
31:10
you're married, but I'm married. I am. And I
31:13
often can predict what my wife is going to say.
31:16
And my wife's autocomplete can often predict
31:20
what my wife is going to say because it knows
31:22
the corpus of her language, right? She sends
31:25
a lot of text messages. And so it knows something
31:27
about the likely statistical distribution of
31:30
the replies to certain messages. But when my
31:35
wife says something that isn't what I predict,
31:38
I can evince a pretty good theory of what's going
31:41
on to make her say something that is out of character.
31:45
And the AI is at sea. And that is the difference
31:48
between understanding and extrapolating. Understanding
31:52
involves a theory and the... promise of big data
31:58
which ai is the latest generation of you know
32:01
15 years ago we were calling this big data and
32:03
now we're calling this ai was theory free inference
32:06
right um the idea behind big data was that you
32:12
could observe by crunching population scale data
32:16
that these molecules taken in this combination
32:20
to people who had these lifestyle factors and
32:23
these pathologies produced a cure or produced
32:29
some kind of improvement in their state without
32:34
ever having to know why, right? That this was
32:37
theory -free inference. And so long as it works,
32:42
you know, it'd be nice to know the theory, but
32:44
so long as it works, you don't necessarily have
32:45
to know the theory. And indeed, there's a lot
32:46
of... medicines we use that are theory -free
32:49
in the sense that we've observed that they work,
32:51
we've double -blind tested them, we've checked
32:54
for side effects and so on, and we use them even
32:56
though we don't understand at all or completely
32:58
how they work. And AI is this idea that we can
33:02
just effectively give up on having a theoretical
33:06
basis for the world and switch to these theory
33:10
-free inferences. And the problem is that it's
33:15
very hard... especially when you're talking about
33:17
unusual circumstances, to unwind false positives
33:20
and detect them when you don't have a theory,
33:23
because you don't really know how to test the
33:25
theory. A theory -free inference is by definition
33:28
not falsifiable. You can't say, if this weren't
33:32
true, then you would expect X or Y to happen
33:37
under these circumstances, because you don't
33:40
have a theory. You don't have a way to know why
33:42
it's true or why it's not true. And so as a way
33:45
of knowing, it's quite limited, which isn't to
33:48
say it's useless. You know, there are lots of
33:51
theory -free things that we have discovered with
33:54
AI that are interesting. And if scientists were
34:00
being told, here is a tool, you have all the
34:04
budget you've ever had before, you have all the
34:06
freedom you've ever had before, and here's a
34:09
tool that does some new multivariant analysis
34:11
that... uh, lets you do stuff that you might
34:14
otherwise have had to hire a programmer for,
34:16
or a statistician, or that you might've had to
34:19
spend several days designing stats, uh, uh, custom
34:22
stats, crunching software for, or algorithms
34:24
for, but this is the thing that will do it for
34:26
you. It will make some mistakes. So you're going
34:28
to have to double check it. You have to be careful
34:30
about how you use it. I bet we would have scientists
34:33
doing all kinds of great things, but that's not
34:35
how it's working, right? What you've got is the
34:37
doge situation, fire scientists, replace them
34:41
with software. make the survivors do more with
34:44
less and expect the same outputs. And I think
34:48
that is very frightening. You know, I'll give
34:50
you an example from my life. So I have an extremely
34:54
treatable form of cancer. And this means that
34:58
I was spending a lot of time with radiologists.
34:59
And in fact, about three weeks ago now, a radiologist
35:03
told me that I'm cancer free, which was very
35:04
good news. And one of the things that we hear
35:09
a lot about is that AI might be able to spot
35:11
solid mass tumors that radiologists can't spot.
35:14
And if there was a sales call at my hospital,
35:17
which is the Kaiser Clinic in Los Angeles, next
35:21
to the Church of Scientology, because that's
35:22
how we roll in LA. And if there was like some
35:26
AI salesman who like, you know, deked out the
35:29
guys from Scientology, trying to get them to
35:31
come in and take a personality test and went
35:33
up to the CEO's office and said, look, here's
35:35
the deal right now. You employ 10 radiologists.
35:38
They do 100 x -rays a day apiece. They cost you
35:41
$300 ,000 each. You're spending $3 million a
35:44
year on radiology and processing 1 ,000 x -rays
35:48
a day. I want you to spend an extra million dollars
35:51
on software. And that's going to sit on the shoulder
35:54
of those radiologists. And once or twice a day,
35:58
it's going to tap them on the shoulder and it's
36:00
going to say, take another look at that x -ray.
36:02
and they're going to do 98 x -rays a day instead
36:05
of 100. So you're probably, in addition to the
36:06
million bucks for the software, going to have
36:08
to hire a part -time radiologist to pick up the
36:10
slack, but you're going to save people's lives.
36:12
I'd be very happy. That, as a cancer patient,
36:16
I love the sound of that. That's not how you
36:19
turn a trillion dollars in annual expenditure
36:21
on $50 billion in revenue into a profit. The
36:25
way that you turn $50 billion of revenue on a
36:27
trillion dollars expenditure into profit is you
36:29
have the AI salesman, deke out the Scientologist
36:32
trying to give them the personality test, go
36:34
into the CEO's office and say, right now you
36:36
spend $3 million on 10 radiologists. Fire nine
36:40
of them. Save $2 .7 million a year. Split that
36:44
between you and Sam Altman. Make the remaining
36:47
radiologists mark the AI's homework. They sign
36:50
their name to the bottom of every radiology report.
36:53
And when someone dies, you blame them. They're
36:55
the accountability sink, right? That's the thing
36:58
to worry about. Now, radiologists right now have
37:01
a ton of labor power. When I talk to radiologists
37:03
about how they're using AI right now, it sounds
37:05
more like the first scenario than the second
37:07
one. But the point of AI, the reason that investors
37:11
think that you can sell AI to people like hospital
37:13
administrators. is that it will erode labor power
37:16
so you can fire nine of the radiologists, right?
37:19
Like that's the reason it's been capitalized
37:22
the way it's been capitalized. That is the politics
37:25
of this artifact as it exists in the contingencies
37:28
that produced it. And while we could produce
37:31
a different set of contingencies and different
37:33
set of politics, that is the politics of this
37:35
version of AI that we're living with right now.
37:38
That all makes sense. But to play devil's advocate
37:40
just for a second, what gets me rattled, are
37:44
some of these whistleblowers, so ex -employees
37:46
who are now warning against the dangers of AI.
37:48
A couple of days ago, I checked out an interview
37:49
with Daniel Cocotelo, who was a researcher at
37:53
OpenAI and apparently refused to sign a non -disclosure
37:56
agreement for $2 million, which would ensure
37:58
that he keeps his mouth shut about their sinister
38:01
affairs or something. He then created AI 2027,
38:04
which I'm heard everybody has heard of. It's
38:07
a detailed scenario of how things might go haywire
38:11
with AI soon. I mean... Listening to the guy
38:14
talk, he just seemed genuinely frightened by
38:17
what he thinks is about to come. And that gets
38:19
me worried. Yeah, look, you know, I don't know
38:22
this guy. I've seen I've seen AI 2027. I don't
38:25
think much of it. And, you know, I've also seen
38:27
the rebuttals from people who don't work for
38:29
AI companies, but do work in AI, like, you know,
38:32
research, the research team at Princeton who
38:34
maintain the AI is a normal technology website
38:37
who wrote a very detailed critique of AI 2027.
38:41
And and I got to say, like, we have a history
38:45
in Silicon Valley in tech. Of guys like going
38:50
into the bathroom, shining a flashlight under
38:53
their chin, looking in the mirror and going,
38:55
and then getting really scared. Right. We had
39:00
it 10 years ago with a bunch of people who used
39:03
to work at Facebook insisting that they had been
39:05
trained to be evil dopamine hacking wizards.
39:08
And that the reason your grampy was a QAnon is
39:10
that they were so good at hacking his dopamine.
39:12
And not that your grampy was a racist forever,
39:15
that you kind of, you know, kept his mouth shut
39:18
about it. And then what Facebook did was like
39:20
trap him in a room full of a bunch of other racists
39:22
who laughed every time he said something racist.
39:24
So now he can't shut the fuck up about it. Right.
39:26
Like that's like a much more parsimonious explanation.
39:29
It doesn't require any wizards. And, you know,
39:31
there is something I think quite exciting. about
39:37
having been an evil dopamine hacking wizard because
39:40
now you can promise that you're a good dopamine
39:42
hacking wizard and the operative word here is
39:44
wizard right like how cool is it to have been
39:47
a wizard and and i'm not saying these guys don't
39:51
think they're wizards i think they might believe
39:54
they're wizards i just don't think they're wizards
39:56
um i mean On another note, this thing is being
40:02
crammed into every single product, whether it
40:04
makes sense or not. At the same time, apart from
40:06
a couple of overzealous LinkedIn boomers, I don't
40:09
know a living soul that would be excited about
40:12
it. Nobody's happy. Half the time, I want to
40:15
take a shower after I've used one of these chatbots.
40:17
And yet it seems we're not really putting up
40:20
much of a fight against this really brutally
40:25
forced adoption. I don't know. I think, you know,
40:28
you see a lot of workers who are getting pissed
40:30
off about it. You see a lot of creative workers
40:32
are getting pissed off about the data. Anti -data
40:33
center movement is gigantic. You know, there's
40:38
a lot of people who are very angry about it.
40:39
I do want to say, though, I know lots of people
40:41
who use AI who I think it was quite credible
40:43
and who say that they got great uses out of it.
40:46
People use it in VFX, not to like render final
40:50
images, but to like automate parts of the process.
40:53
The best programmer I know is a guy called Patrick
40:56
Ball, who founded an NGO called the Human Rights
41:00
Data Analysis Group, HRDAG .org. And HRDAG, they
41:04
specialize in using fragmentary evidence from
41:08
war crimes, civil wars. genocides, and they use
41:13
rigorous statistical methods to build out a picture
41:16
of what aspects are probably missing from the
41:18
data by interpolating the data. And then they
41:20
testify at human rights tribunals, truth and
41:23
reconciliation. They worked on Milosevic, and
41:25
they worked in Rios Monte, and East Timor, South
41:29
Africa, the Colombian Civil War. Patrick's, as
41:32
I say, one of the best programmers I know, brilliant
41:34
statistician, brilliant programmer. He tells
41:36
me he's using a whole bunch of AI agents to write
41:39
the best code he's ever written. Right. But Patrick
41:42
also has more discernment than any programmer
41:45
I've ever met. So, you know, if I was ever going
41:48
to trust a programmer to know which parts of
41:49
the job to automate and how to double check to
41:52
make sure that when you automated it, you hadn't
41:55
accidentally produced a bunch of tech debt. Patrick's
41:59
the guy I would trust, you know. And so, again,
42:03
this is one of those things where you have labor
42:05
driven adoption. Look, I use I use a chatbot
42:09
running on my computer, a local open source model
42:11
to check my blog post for typos in the morning
42:14
because I publish my blog on all these different
42:16
platforms. I publish it. as threads on all the
42:19
different social media platforms. I publish it
42:21
on Tumblr and Medium and I put it on a newsletter
42:23
and I put it on the web. And when I find a typo,
42:26
my readers really want me to know about it. They're
42:28
very invested in my knowing about my typos. And
42:30
so I have to go fix it or I'll just get like
42:32
a million emails. And once I published it to
42:34
10 places, fixing just like a period that's gone
42:36
astray is like 15 minutes work. And so if there's
42:39
six typos, that's like an hour and a half that
42:41
I lose, right? So I drop it into this local chatbot.
42:44
I type find typos. and finds all kinds of typos
42:47
right it's just like and and many of them aren't
42:50
typos and i ignore those because i have discernment
42:53
i've been telling search engines that they were
42:56
or not search engine spell checkers that they
42:58
were wrong for 30 years uh i've published 30
43:01
books more than 30 books with major publishers
43:03
i've been telling copy editors and proofreaders
43:06
that they were wrong for 25 years right and i'm
43:09
perfectly capable of telling a chatbot it's wrong
43:11
too The difference, of course, is the chatbot
43:13
never learns. Whereas the proofer, when you say
43:16
that's not a typo, that's intentional, doesn't
43:18
send you back the manuscript next time with those
43:20
errors in it. It's not perfect by any means,
43:23
but like I was never going to hire a proofer
43:25
to read my blog post in the morning in the 10
43:27
minutes after I write it before I publish it.
43:29
It's just ridiculous. So yeah, I'm very happy
43:32
to have it. It's fine. It is absolutely fine.
43:35
It is a normal technology. My first word processor.
43:40
was a program printed in a magazine that I bought
43:43
at the corner store and took home and typed into
43:45
my Apple II Plus. I was 10 years old. It was
43:48
1981. I've used a lot of word processors since.
43:52
A lot of them have had new features. A lot of
43:55
those new features I thought were stupid, and
43:57
when other people use them, I think it made their
43:58
writing worse. Some of them were really useful,
44:01
and I use them all the time. I've never seen
44:04
a feature for a word processor where I said,
44:06
that's so cool, we should bet the world's economy
44:08
on it. And by the way, we should probably also
44:10
put all the writers in a wood chipper now that
44:12
it exists, right? And like, that's the problem,
44:16
right? Is we are treating what cashes out to
44:19
plugins for common software that are sometimes
44:22
useful and sometimes not. And when workers drive,
44:25
the automation can produce things that make your
44:28
life materially better. And we're treating them
44:30
as genies and gods and demons, as the end of
44:34
the world or as the beginning of a new one. And
44:37
it's not that. If AI destroys the world, it's
44:40
going to be because the bubble has burst. It's
44:42
not going to be because the software was so cool
44:46
that it woke up and turned us into paperclips.
44:51
Maybe this is a mad thought that's completely
44:53
divorced from reality, but is it possible we
44:55
just get... bored by it all? I mean, how many
44:57
more AI generated videos of Tom Cruise fighting
45:00
a large Pokemon on some Chinese roof can I watch
45:03
on X before I just want to retreat to a prehistoric
45:06
cave forever? I mean, I can't be the only person
45:08
that's already tired of these smart ass chatbots.
45:10
Come on. Yeah. And I got to run. So this is maybe
45:12
a good place to end the conversation on. I do
45:15
think that like a lot of that AI stuff is getting
45:17
washed really quick. And, you know, this has
45:19
always been one of my critiques about a certain
45:21
kind of technological doomsaying that even where
45:24
you have a thing. that works really well at the
45:27
start, right? Someone's found a little cognitive
45:29
blind spot. Someone's found a thing that, you
45:32
know, makes people curious or catches their attention,
45:35
that unless there's some substance to it, that
45:39
it very quickly becomes something that you're
45:43
inured to, right? You know, do you remember when
45:46
Upworthy headlines were invented and it was like,
45:48
11 facts about socks, you'll never believe number
45:51
three. And I clicked through like half a dozen
45:54
of those headlines. And then I realized that
45:56
like number three just wasn't that interesting
45:58
and never would be. And I never clicked through
46:00
one of those headlines again. Now, that's not
46:02
to say that that form of headline in the hands
46:05
of someone who's a good writer, who knows how
46:07
to infuse it with actual communication, who's
46:10
creative, who's smart, and who has something
46:12
to say. Because, you know, the problem with those
46:14
upworthy headlines is they were always on articles
46:16
that were stupid. right? The best thing about
46:19
them was the headline. But if you have something
46:20
smart to say, there are probably aspects of those
46:23
techniques that will work to capture people's
46:25
attention and hold it because you're writing
46:28
about something real and substantive. And I believe
46:31
that we will have AI generated text and images
46:34
and so on, not in vast quantities, not fucking
46:38
emails, routine business emails that have been
46:41
inflated three bullet points into five paragraphs.
46:44
you know, florid bullshit, but like we will have
46:47
AI generated material in our lives that will
46:50
be being used in the background, whatever, you
46:54
know, to, to help you get stuff done, uh, and
46:58
to, you know, make things better. And some of
47:02
it will be stupid and some of it will be fine.
47:04
And most of the people who are making it now
47:06
will have moved on to something else because
47:07
they're, they have nothing to say. They're just,
47:09
they're just fiddling around with a new web toy.
47:13
And in that regard, AI is a normal technology
47:16
because I've just described the trajectory of
47:19
many technologies that we've had in the last
47:22
couple of decades. That's how it went with BuzzFeed
47:25
quizzes, right? That's how it's going to go with
47:27
other web toys too. Corey, always a pleasure.
47:32
Thank you so much. Nice to chat with you. Thank
47:34
you so much for listening. Please subscribe on
47:35
YouTube, follow and rate on Spotify and Apple
47:37
podcasts. You would help me a bunch with that.
47:40
Also feel free to support me on Patreon. This
47:42
is a one person podcast. I do everything by my
47:44
little old self. It's a lot of work, especially
47:46
with no team, no ads and no real marketing, but
47:48
you can help by hopping on Patreon and finding
47:50
the smart cookies podcast on there and giving
47:54
me your money. Of course, super hard for small
47:55
creators to get through. So your support. would
47:58
mean a lot. And then last, but certainly not
48:01
least, I have to thank my existing patrons. Without
48:03
you, there is no Smart Cookies podcast, which
48:05
makes you not only absolute legends, but also,
48:09
in effect, actual producers of this show. So
48:12
eternal thanks go to Lennox, Dumi Holdis, Čertomir
48:18
Perherić, Hanna Zirman, Dermastija Ponapasion,
48:20
Charlie Ballbeck, Zala Veljkovar, Matija Valingarsluga,
48:23
Katja Glinšek, Pač Mačkin. You guys are the best.
48:48
I'm deeply touched. So thank you.