November 20, 2025 • 51 mins
In many respects, AI technology is already mind-blowing, and can perform many tasks far better than the average person. And yet by and large, its impact has been hard to detect. We haven't seen some huge labor displacement, for example. There's nothing dramatic yet happening in the productivity data. So when will the impact really start to be felt? On this episode, we speak with Tyler Cowen, a professor at George Mason University and the co-author of the famed econ blog Marginal Revolution. He's also the host of the Conversations with Tyler podcast. We talk about when we'll really start feeling AIs impact, as well as other topics, like food, music, and the general state of discourse.
Subscribe to the Odd Lots NewsletterJoin the conversation: discord.gg/oddlots
See omnystudio.com/listener for privacy information.
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:02):
Bloomberg Audio Studios, Podcasts, radio News.
Speaker 2 (00:18):
Hello and welcome to another episode of the Odd Lots podcast.
Speaker 3 (00:21):
I'm Jolle Wisenthal and I'm Tracy Alloway.
Speaker 2 (00:25):
You know, not a novel observation here, but I do
think that if, like in early twenty twenty two or
late twenty twenty one, you had someone had revealed to
you all of the amazing things that we could do
with AI these days, I think you would have expected
that either the broader economy or society would be more
different than it has been. Like I am large, you know,
(00:49):
I think most people's jobs are done roughly the same way.
Sort of society still seems to operate, although a little
maybe a little bit worse every day. I don't know,
But like I just don't think like it's mind blowing
technology and get by and large, like it hasn't had
the economic disruption that I think many people would have guessed.
Speaker 3 (01:07):
Maybe I'm cynical on the subject, but I always say,
never underestimate the human capacity for stasis, I guess, and
making things much more difficult than they actually need to be,
and putting up you know, bureaucratic barriers, regulatory barriers and
things like that. So it's not it's not that surprising
to me, but it is true. You have a lot
of economists out there who think there was going to
(01:27):
be this massive productivity boost, right well, it's certainly.
Speaker 2 (01:30):
Tech people like I don't know, like economists. You know,
they're like, oh, there were some we're going to have
a productivity boom. We're raising our estimates from two percent
too and a quarter percent, and then you have all
things relative to and then you know, we talked to
Kathy Wood one time, and I think she what did
she predict twenty percent real GDP growth for twenty years?
Or you certainly have these people in Silicon Valley deflationary boom,
(01:54):
post scarcity any any minute, right now, some real gaps
I would say, between how lot of economists think about
some of these things and the numbers, then economists would
be comfortable using versus maybe literally.
Speaker 3 (02:07):
Everyone else like Kathy would certainly many others. Yeah, you
know what I think about is how different would your
career in my career have been had AI existed in
like two thousand and eight, two thousand and six when
we were starting to blog.
Speaker 2 (02:22):
Basically, it's a really good question, and I don't know
the answer because part of me thinks, well, you know,
that whole path that was like defined my career like
would not have been there, would not have existed or no. Maybe,
But on the other hand, maybe in two thousand and
six I would have been like just like I was
super early into blogging, super early into experimenting with AI news,
(02:45):
and it would look different, but maybe I would have
rowed some different.
Speaker 3 (02:48):
Different but the same.
Speaker 2 (02:49):
Yeah, different but the same, you know, Like I think
it's a it's pretty hard to tell.
Speaker 3 (02:53):
One thing I know is like back then we were
writing to optimize Google results, right, that was basic or know,
like I know social media was disseminating stuff as well,
but like Google was still the main platform for a
lot of stuff. I'm sure if we were doing it now,
we would be optimizing for chat, GPT or perplexity or
someone like that to actually pick up the content. I mean,
(03:16):
I think that's the audience.
Speaker 2 (03:17):
Yeah it is. And I think like any especially any
like commercial publisher, right that's trying to do a scale.
I think like a sort of niche voice expert could
still have like their audience that comes to them directly,
of course, but like it's scale scale for sure every
publisher is trying to figure this out. Anyway, We really
do have the perfect guest, someone we should have had
on the podcast years and years ago. It's almost surprising
(03:39):
that it's just the first time we've had him on.
It is crazy, It is crazy, but it is someone
who I think is like at the intersection of everything
that we're talking about. Because he's an economist. He knows
the tech really well. The tech people know him, the
tech people love him. He may even be one of
the preferred economists for this. I feel he's like the
economist that all people in the AI world want to
(04:01):
get his take. A long time blogger, someone I, you know,
like both of us sort of gave up on mike blogging.
Although we have our newsletter. It's like close enough, but
it's different than blogging. Someone who's stuck with the medium
for a long time. One of the true original econ
bloggers that I've been reading for over twenty years. We're
going to be speaking with Tyler Cowen. Here's the host
of the Conversations with Tyler podcast. He also, obviously is
(04:22):
one of the two bloggers that the famous Marginal Revolution blog.
Economics professor gmu appreciator of ethnic foods all around the
DC Northern Virginia, Maryland area. Someone known on the internet
for a long time, Tyler, thank you so much for
coming on odd Laws. Really thrilled to have you here.
Speaker 4 (04:44):
Hello, happy to be here.
Speaker 2 (04:45):
Amazing. What do you think about my initial assessment? How
fair do you think it is that, Like, had you
known in twenty twenty one how powerful these tools would be,
maybe we'd be a bit surprised that by and large,
like busines, and this seems to more or less run
the same, I'm.
Speaker 4 (05:03):
Not surprised at all. So what I see right now
is people using AI as an add on to their
pre existing work routines. Oh, you need to write a memo,
you ask AI how to do it. You need to
write a column, you ask AI to proofread it or
fact check it. And that works great, But those are
marginal gains. What we really need to see a major
impact is new organizations built around AI, and those will
(05:27):
be startups. They will come only slowly. It will take
twenty or more years before they really transform the economy.
And in the meantime, it's a whole bunch of add ons,
which are fun and fine, but that's why I think
it's slow.
Speaker 2 (05:40):
What can you tell us about the history of technology
such that legacy organizations that existed prior to the invention
of maybe potentially revolutionary new technology have a hard time
massively changing their workflows.
Speaker 4 (05:58):
Well, you can take even very simple examples. So Toyota
starts competing with General Motors in the nineteen seventies, General
Motors is paralyzed. It cannot really come back and adopt
the new and superior Toyota methods, and those are not
really that different compared to say AI. So there's just plenty,
plenty examples. Old mainstream media could not cope very well
(06:20):
with the Internet. There are exceptions like The New York Times.
Odd Lot's podcast would be another thank you, But it's
the norm. So we're seeing it again with AI. And
again you need a complete turnover of who and what
is doing business for it to really matter on a
big scale.
Speaker 3 (06:37):
If you need that complete turnover, and you need some
time for AI to become fully embedded in a business
model or for a business model to form around AI,
what industry or what part of the economy would you
expect it, I guess to show up first in that
sort of revolutionary way.
Speaker 4 (06:56):
Well, we have obvious data on this, and it's programming.
You will hear people who do programming claim that say
eighty percent of the work is now done by AIS.
I suspect that's an overstatement, but there's no doubt at
all that there's simply a lot of programming already done
by AIS. When you have low fixed costs, a competitive sector,
immediate feedback, you know, the revenue has to flow programming.
(07:21):
And also New York City finance, there's been quantston finance
for a long time. Those quants are now you could say,
more AI equipped than they used to be, and those
areas already are being revolutionized.
Speaker 2 (07:33):
I've heard them and got some pitches to do episodes,
which we should do at some point. But I've heard
about like some law firms that are being like new
AI law firms where there are lawyers, et cetera. But
from the very ground up, the idea is perhaps there
is some way to just get way more productivity if
(07:54):
they start from the very beginning with some combination of
lawyers plus AI models. It seems like that could be
the kind of thing where maybe the legacy law firms
are seeing some productivity gains from AI. There's probably some
evidence you could find but that a new one with
a totally different approach could deliver that big productivity boost
(08:14):
that actually ends up changing the industry.
Speaker 4 (08:17):
It's already the case, say, mid tier associates are much
less needed. But there's one big problem with law in particular,
and that is the way large language models work. Now
you have to send your queries somewhere else. You can't
just own and control and hold the whole thing on
your hard drive. Now. I think within a few years
time that will be very different. But until then, major
(08:38):
law firms are extremely skittish about just typing in their
questions and sending it, you know, to San Francisco. I
don't actually think there's a risk, but when you think
about how fiduciary responsibility works, they just don't want to
do it.
Speaker 3 (09:07):
Talk more about I guess privacy concerns and regulation, because
this seems to be an area where if you are
in a heavily regulated industry, or if you're just in
an industry that tends to be full of paranoid people
like lawyers, it does seem like there's going to be
a natural tendency to be very very cautious. When it
comes to sharing data with AI, you're going to be
(09:28):
worried about actual data ownership the queries that you're sending
to San Francisco, as you say, are those industries just
inevitably going to be slow to adapt.
Speaker 4 (09:37):
They'll be slow to adapt again until the point where
they just control their own model and they hold it
within the firm and no one else really can get
at it. So what you need is cheaper models where
a law firm can afford to have its own model.
And I think that's a few years away. It's not
a very long time away, but it won't come in
(09:59):
six months. Sam Altman I just did a podcast with
him and he said a privacy problem is AI queries
or subject to subpoena, and he thinks they should have
as much protection as say, your conversation with your lawyer
or your doctor or your therapist. I think that's a
good idea, but that hasn't happened yet, and until it happens,
(10:19):
or you get the whole thing on your own hard drive,
progress in law is going to be slow. But once
the progress comes, that's one of the areas where I
think AI has the most promise it's just very very
good at mastering a large corpus of text and organizing
it for you.
Speaker 2 (10:35):
It is interesting, isn't it. Where it's like, Okay, if
you are my lawyer, you and I could have a
conversation and it would be not it would not be uh.
Speaker 3 (10:44):
It'd be privileged.
Speaker 2 (10:45):
Yeah, it would be privileged. Or you know, one can
have conversations with anyone, and as long as you're doing
it on the phone, or if you're doing it person
to person, it's much more difficult to get that, you know,
in discover or I always think about this with you know,
public public records of employees that sometimes you can get
(11:06):
their emails through Floyers, et cetera, but you can't get
the cons content of conversations. It does feel I hadn't
really thought about this dynamic though, that when you're using
an AI it is sort of like a conversation and
yet sort of from an evidentiary basis, it would be
much more like an email.
Speaker 4 (11:25):
I think when it comes to medical issues, there are
many more people willing to share their data. Not everyone
by any means. Some medical conditions are secretor people just
don't want others to know. But I see many, many
people I know typing in all kinds of things about
their medical history to say GPT five and getting what
are on the whole very good answers. It's like medical
(11:45):
diagnosis for free spreading now to the whole world. A
lot of countries where people just don't even have access
to good doctors at all, and I think that will
be important more quickly than the law innovations.
Speaker 3 (11:59):
Just as a thought experiment, what does all of this
mean for insurers? Because I kind of think, you know,
I think about a bunch of people typing in their
medical information. I think about basically the explosion in data
that we have nowadays, and it seems like the insurance
industry would be one place that would really benefit from
all this trove of additional information if they could access
it well.
Speaker 4 (12:21):
This is one of my worries about AI in general.
I'm quite positive on what's happening, But as insurers get
better and better information on their customers, this is just
through big data more generally, it doesn't have to be
current large language models. They know exactly how to write
the risk and how to price the premium, and in
a sense for the buyer, it's not insurance anymore. So,
(12:41):
if we know your house is going to burn down
with high probability and you have to pay the super
high premium, you don't really have the benefits of the insurance.
So some insurance markets might unravel if through big data
the insurers learn too much about what's likely to happen.
Speaker 2 (12:57):
Economists seem to be very consistent about the effects of
technology on labor demand, which is that in the end
it washes out right, some people, Okay, there's a disruption,
but I'm gonna save money because I use the I.
But that means I have more spending power, and then
I'm gonna buy something else that I wouldn't have bought
had it not been for paying wages, and then that'll
(13:20):
create demand for labor elsewhere. And so in the end,
the idea that you could really have tech driven unemployment
at scale that is not transitory or not temporary is hard.
A lot of economists seem to be intuitively skeptical of
this idea. Whereas you have people in the ifields, fifty
percent of the people aren't gonna have jobs we need
(13:41):
ubi otherwise there's gonna be a permanent underclass. Could there
be something different about AI such that it doesn't have
the same labor market effects that past technologies have had.
Speaker 4 (13:52):
I would say, you understand me. Well, So the energy sector,
there's going to be a lot of new jobs taking
care of older people. I think as AI produces more potential,
medical innovations will need to test them, So the biomedical
sector testing clinical trials, there'll be a lot new jobs.
I'm not worried about mass unemployment, and most economists are not,
(14:14):
and I agree with their perspectives, which I think you
outlined pretty clearly.
Speaker 2 (14:18):
What do you say, though, because I have a feeling
that when you're out in San Francisco, they don't see
it that way, and they talk a lot, and some
of them are more in you know, UBI talk and
permanent underclass talk and all of this stuff. Do they
see your perspective when you make this case? What do
they say or what are they What are a lot
of them seem to be missing about the logic that
(14:39):
you spell out.
Speaker 4 (14:40):
Well, I think more and more they're coming around to
the economist's point of view. So Andre Carpathi, who was
at open AI, you know, in the most important years,
he just did a podcast saying he thinks adjustment will
be slow. Things will be fine, will grow at two
points something percent. There won't be mass unemployment, and you
wouldn't I have heard that, say two years ago. But
(15:02):
I think as people see the models rolling out, and
as you mentioned well, the real world impact it is
stretched out in time, right, it's not all immediate. Earlier
on people had more the sense that AI was a
kind of god box that you just talked to it
and it can magically do anything and convert that into
results in the real world. But if you think about
(15:22):
your actual job, even if it's a highly intellectual job,
so much of what you do is the interaction between
your intellect, your physical presence, your interactions with others, your travel,
many other things. And until we get to some far,
far off world where the robots are perfect copies of you,
which I don't think ever will come, jobs will be fine,
(15:43):
but they will change a lot, and I'm actually worried
about who will be the biggest losers. I think poor
people will do great, the very wealthy will do well,
but people who are sort of upper upper middle class
will find this automatic ticket to a law or consulting
job that assured they would be upper upper middle class
(16:03):
for the rest of their lives. I think a lot
of that is going away already.
Speaker 3 (16:08):
I think also no one would have expected plastic surgery,
I guess, to be a beneficiary of the AI revolution.
But if you think that what's going to matter in
the future is like your personal presence and your network
of social contacts, then I guess we should all be
working on on our looks Maxing, Yeah, absolutely right, charisma. Okay, noted,
(16:29):
everyone work on their respective charisma. One thing I was
wondering is the impact of AI on public finances. I
am not very good at tax policy. Joe knows this
because I've complained about taxes to him repeatedly. But I'm thinking,
but if you're thinking about where the value add of
(16:49):
AI actually shows up in the economy, So you know,
presumably you got a productivity boost. We're not entirely sure
how much that's going to be. But where does that
additional output actually show up in terms of revenue for governments?
How is that collected? And how would you expect the
distribution to vary across the world.
Speaker 4 (17:07):
Well, I think in the United States medium term, they'll
just be much much more. Healthcare will be new drugs,
new medical devices, will have to test all these things,
will have to produce them, and that will be It
was already the case that those sectors were growing, but
that growth will be accelerated. So that's where I think, God,
they'll be the biggest difference longer term, and people will
(17:29):
live longer because we'll fix at least partially various diseases
and maladies. So if you live to be ninety four
across a lifetime, you spend way more on healthcare than
if you live to be seventy seven. And that's yet
further growth for the healthcare sector. But some things like
medical diagnosis that's already very cheap, like you know, a
(17:50):
good large language model probably outperforms your current doctor at
least if you type in what's wrong with you properly.
Speaker 3 (17:57):
But does that additional productivity or or the output generated
by AI, does that actually show up an additional taxation
for the government.
Speaker 4 (18:06):
Well, the healthcare sector generates an enormous amount of taxation revenue.
I do think we'll have some sectors that maybe just
become free in the same way that Wikipedia is free.
So I could imagine say ten or twenty percent of
the music sector is music you create at home using
your own AI and it's a customized song for you,
(18:26):
and maybe you paid a subscription for the service. But
rather than spending more money on Spotify or a streaming service,
you just build the music and that's a partial substitute
for some human created music. I don't think human created
music will go away at all. People want to enjoy
the human touch, the feeling that you're a fan of Taller,
Swift or whatever. But there's going to be a lot
(18:48):
of AI generated music and art and many other areas,
and some of it will be free. But that's not
a problem from a revenue standpoint. So instead of spending
money on safe buying a picture, you create one digitally
at home with your AI. You'll spend that money on
something else.
Speaker 3 (19:06):
I got to ask, now, are you a swiftie.
Speaker 4 (19:09):
No, it's worrying for me. It's a little too predictable,
so I have to say I'm not, but I'm glad
other people are. Let's put it that way.
Speaker 2 (19:19):
Do you have a theory let's talk, let's just talk
about music. Do you have like a theory of the
swift phenomenon? Like, is there a reason? What is it?
Because it's just so, doesn't I don't know what's your
what's your meta take on the tailor, the swifties and
culture and society.
Speaker 4 (19:38):
Well, it's super polished and because of the Internet, the
very biggest of celebrities can be much bigger than before,
and someone will fill a few of those slots. But
I think also how she presents herself. She has the
guise of being attractive without feeling threatening to other women. Uh,
and there's something all American about her and quite generic.
(20:00):
She doesn't rule out the fandom of that many people
m hm. And she's the one who's filled that slot.
She's been brilliant at managing her career and seems to
just stick at it and has an incredible work ethic.
People say the shows are incredible. As more and more
life goes online, who can give a good show? It's
the charisma and looks. Point Well, she seems a pleasant
(20:21):
that I've never been to one, but I hear plenty
of reports. And you put all that together and she's,
you know, the megastar of the music world.
Speaker 2 (20:29):
Do you think like culture, Like there's this popular idea
that culture is sort of dead, and I do think
that is probably overstated, But you know, you look at
movies and people have people have observed this for a
long time. It's just rehashes of franchises that have been
around for thirty years. And I think if you look
at Spotify's streaming, there's still this overwhelming tyranny of the
(20:52):
boomer rock et cetera, and this feeling that culture in
many respects is that's rehash that is very hard for
new things to to break out. I mean, Taylor Swift
at this point is decades old phenomenon. Is that real
in your view? Or is that just people in the
pundit class who've been lazy and not discovering new things
(21:13):
and not actually putting in the effort because they're not
young anymore and they're not going out and they say
nobody listens to new music.
Speaker 3 (21:19):
Joe, are you just talking about yourself?
Speaker 2 (21:20):
I have talked about I'm trying. I'm doing a little
introspection here. Is this just me because I don't go
to shows like I did when I was twenty or
is something changed?
Speaker 4 (21:28):
But a lot of it is the pundit class. So
you look at movies, I think it's perfectly correct to
say the most popular movies today are pretty dreadful, and
it used to be the case that the most popular
movies were God, The Godfather in Star Wars. Yeah, that's
a big change. But if you look at movie making
around the world and in a given year list say
the twenty five best movies from all places, which may
(21:51):
not even come to your multiplex every year, you have
an incredible list. I don't think they're worse than the
movies of earlier times. I do think mainstream hollywo is
much worse. So in many areas you just have quality
moving more into unises.
Speaker 2 (22:06):
By the way, I just want to say, the first
time I ever encountered your work, prior to even having
stumbled across Marginal Revolution, was at the bookstore in Austin
finding a copy of In Praise of Commercial Culture, and
I just feel like that book is held up so well.
I mean in the specific sense that there is so
(22:26):
much mass culture these days, whether it's high end Netflix
TV shows, et cetera, whether it's A twenty four films,
whether it's Beyonce or Taylor Swift or some of these
other big names who are simultaneously incredibly popular. And I
get that you're not as swifty, neither mine anymore, but
people take this yeah, I liked their country for once
(22:49):
she left country. But like where people take these popular
culture things extremely seriously as art and don't dismiss these
outputs as sort of being trash.
Speaker 4 (23:02):
Right now, we're in a golden age for country and
western music and also horror movies. Neither is really like
my taste in particular, but it's easy to see what
has gotten worse, and especially for critics, harder to see
what has been getting better.
Speaker 3 (23:17):
That's my sweet spot. Country and horror is perfect for me.
But just on the culture point, I mean, I think
the lack of culture argument, the one I hear the
most is it's a lack of shared culture.
Speaker 4 (23:28):
Right.
Speaker 3 (23:28):
So you do have like some giant monolists like a
Swift or a Beyonce that everyone knows about, and you know,
they do have these large audiences, but broadly we're not
all experiencing the same media that we used to, right, Like,
no one is gathering around the TV to watch the
finale of you know, some show that airs like once
a week and has been going on for five years.
Speaker 2 (23:49):
The exception is NFL football, which I don't really get it,
but yeah, but it's not cultural.
Speaker 3 (23:53):
Sports is outside of my experience.
Speaker 4 (23:56):
And the Super Bowl is cultural, right, it's a cultural event.
You care about the advertisements, But the biggest of YouTube stars,
which again I would say is not personally my thing,
but they can have bigger audiences than those older TV shows. Right.
Speaker 3 (24:10):
Well, So what I'm getting at is it does feel
like nowadays there's an ability because of tech, to serve
up very specific content and like niche content in streams.
And the analogy that I like to use is, you know,
everyone knows if you download Netflix for the first time,
the first movie you watch is incredibly important because whatever
(24:33):
you watch, you know, if it's a rom com, you're
going to be served up Kate Hudson films for the
rest of your life. Right, Like, the algorithm looks at
what you're watching and then it serves up that additional content.
What does that mean for society? The idea that you
have people you know, basically funneled into smaller and smaller
streams in some respect, well.
Speaker 4 (24:54):
A lot of the Netflix algorithm it just directs you
to slop. True, people have always wanted slop. Like people
listen to music way back when it was quite common,
or they would just list in the top forty, which
in some years was very good, but often was pretty terrible,
even in the nineteen sixties. So what you can do
today is basically watch not any movie out there, but
(25:18):
you have movie You can still buy DVDs and Blu rays.
You have access to more cinema today than you ever have.
So people will sort themselves. And I think it's from
the point of view of cultural consumption. I don't think
there's ever been a better time to be alive than
right now. Like, well, a lot of people abuse that
(25:39):
and go for slop. Of course that's sad, but it's
hardly new.
Speaker 2 (25:43):
So changing gears a little bit. Tracy and I write
almost every day because we have a daily newsletter that
forces us to and I really like having that because
I don't know if I would write every day if
I did not have that obligation to deliver something in
people's inbox that they pay for as partler Bloomberg dot
Com subscription. I love writing, but I don't know if
I would do it every day if I didn't have
(26:05):
this sort of requirement. I might just tweet. How do
you You've been blogging for over twenty years, how do
you avoid the temptation to just fire off all your
ideas via tweet and actually commit to the blog.
Speaker 4 (26:19):
I'm never tempted to do that. I like to think
things out.
Speaker 2 (26:22):
I don't do.
Speaker 4 (26:23):
I just don't more when I write properly. I've actually
blogged every single day for over twenty two years.
Speaker 2 (26:29):
That's amazing. It's amazing. Most people gave up, and so
what's the difference.
Speaker 4 (26:34):
I don't feel it requires any discipline. For me. The
discipline is not writing more like I have to restrain myself.
So I guess I'm just weird. I don't think I
have any neat little trick or formula. It's one of
these niches that you can do now that you couldn't
do before. And I found my niche, as have the
two of you.
Speaker 3 (26:53):
Here's a slightly different question playing on that theme, but
going back to the intro, How different do you think
you're blogging career would have been had chat bots existed,
you know, ten or twenty years ago when you were
starting out.
Speaker 4 (27:07):
I don't think we know yet. My intuition is that
people still want to read human writers simply because they're human,
and if the bot is as good as you, most
of the world doesn't care. But that has not truly
been tested yet. I think we'll see in the next
two years, but that's what I'm expecting. Just I think.
I think in music, there'll be plenty of AI music.
(27:29):
It might be say ten or twenty percent of the
music sector, but listeners will still want that human to
human connection.
Speaker 2 (27:51):
Do you think you know, when Twitter came out, they
called it a micro blogging site, as if it were
just blogging but on a shorter basis. But I think
it's fundamentally different. You know, in the early, the glory
days of blogs, which we'll be talking about forever when
we're all very old people, how good it was, you know.
I think there was this sort of spirit of you know,
(28:11):
liberal linking with each other and idea exploration, whereas Twitter
strikes me as much more conflictual and one upmanship and
so forth. Do you think there are like fundamental I
don't know if political is the right word, but like
new communication paradigms like sort of have their own terroir,
so to speak, in terms of the impulse towards collaboration
(28:34):
or conflict, etc. And does that change society?
Speaker 4 (28:38):
Yeah, I still like blogging, and I'm sad people have
moved away from it Twitter. To me, it seems too
meme heavy, and meme heavy media have more potential for racism,
which of course is a big negative. And I see
so many people who are driven crazy by being on Twitter.
Whether it's because they're writing on it or reading it,
I'm not sure, maybe both. I don't want to name names,
(28:59):
but it's a lot of people, and I bet you
see the same ones that I do.
Speaker 3 (29:03):
Also very sexist nowadays, I would just add in unappreciated
ways in many ways speaking of sexism, the impact of
AI on economics. Talk about that economic institutions, you know,
famous for modeling, spend a lot of time with numbers
and things like that. Is that all just going to
be replaced by AI?
Speaker 4 (29:23):
Not all? So. I think what human economists will do
is put more and more time into gathering data and
feeding it to the AIS. The returns to doing that
will be very high. But the actual econometrics, statistics humans
will maybe set up part of the problem, but the hard, boring,
routine work will be done by the machines. As to
(29:43):
some extent it was already the case, and this will
be a way to make a lot of progress rather
quickly do the.
Speaker 3 (29:49):
Actual economic statistics or data points that economists collect. Do
some of those need to be changed or thought of
differently in light of the AI era.
Speaker 4 (30:00):
Well, I don't know. Eventually they will need to be,
but I would say any period of radical change in history,
your statistics are less useful. It's not that the people
creating the statistics are making some mistake. You just cannot
capture every way the world is changing, and index number
comparisons require the basket of goods be relatively close to constant,
(30:23):
and at some point that doesn't hold anymore, and we'll
be faced with that. We'll deal with it. I would
say the current statistics we have, they're more underrated than overrated,
so they're actually pretty good. I'll be glad when we
get them back again.
Speaker 2 (30:38):
There's another thing. The tech people you talk to, they
must think like GDP is terrible. Doesn't capture all this stuff,
all this value that we can't price. I'm sure you've
had conversations explaining to many AI workers that GDP is
not the worst statistic in the world. I want to
go back though, to them. So you know you mentioned
like on Twitter, right, you say something someone so it's
(31:00):
a meme. They dunk on you, they make fun of you,
they whatever. Your chatbot won't do that. Like if I'm
having a conversation with Chad GBT, it's never going to
respond to me with the meme sort of indicating one
day of my ad that I'm.
Speaker 3 (31:14):
A moral you'll learn your language, Joe.
Speaker 2 (31:16):
Yeah, But if anything, it is too obsequious, right, I mean,
the issue is like the online world has become this
very like sort of competitive, conflictual world, and then I
go to the chatbot and my complaint is literally the opposite.
It doesn't challenge me enough. It's too obsequious. Every question
I ask, it's a great question. Sometimes I wish it
would call me a moron a little bit more. But
what does it change about the world? And we know
(31:38):
that like many people's brains have been broken by social
media that probably has downstream effects on how our politics
operates these days. What does it do to the world
if we started inhabiting these chat environments where they're just
very sort of polite, and every time you say something,
it says, yes, great thought, Tyler, great thought, Joe, would
you like to expand it? You asked the perfect question,
(32:00):
what do you see that having sort of second order
effects on how society operates.
Speaker 4 (32:05):
Well, that's the four to zero model that does that.
The newer models like Claude four point five and GDP five,
they're more objective, and that's better.
Speaker 2 (32:13):
But they never make fun of you. They never will
like say, you're a moron? How could you possibly how
could you possibly ask such a dumb question? You're obviously
so out of touch for having the ax ask this.
This is like a very in some respects. This is
a very positive change from many of the conversations that
I've had from typing into a computer.
Speaker 4 (32:34):
Oh it's great. I think people should be nicer to
each other. And I think they're the most objective media
source the human race ever has had. If you're ask
good about, say vaccines or conspiracy theories, it basically gives
you the right answers.
Speaker 3 (32:48):
Well, one thing that they don't do, and I mean
I do think they can be trained to be mean
to you and to insult you to a certain degree,
and you can.
Speaker 2 (32:56):
Could bring but I never account of them.
Speaker 4 (32:58):
Try market that more readily on PAP right.
Speaker 3 (33:02):
So we spoke to the chief business officer of Perplexity,
Dmitri Shevalenko. We spoke to him recently, and he was
saying that one thing chat models can't do is express
a natural curiosity, which I thought was kind of weird
coming from him, because perplexity is the only model I
know that actually throws out those additional questions if you
query it and then it comes up with would you
(33:23):
like to have more information on this point or are
you thinking about this now? But there does seem to
be an element of creativity perhaps that is lost in
some of these lms. How much does that change things
in media? The idea that you know, the models are
going to spit out something that's sort of predestined in
many ways.
Speaker 4 (33:44):
Well, that's what most humans do, to be clear, But
it's now the case that on a regular basis, the
models say can prove new theorems and math or discover
new potential drugs. And keep in mind, you know, a
year ago these things thought the word strawberry had too
and now they're winning gold medals and math Olympias. So
a year or two from now, maybe we don't know
(34:06):
how much better they'll be, but I don't think they're
gonna have any problems being creative, certainly more creative than
humans on average.
Speaker 3 (34:14):
So now I have to ask. Since we're talking about
being mean or nice to the models and them being
mean or nice to you, do you say please and
thank you in your LLM queries?
Speaker 4 (34:23):
You know? I used to, and then Sam Altman said, well,
it costs us just a little bit of money because
of the extra tokens. And then I thought, I'll hold
off on this. But I have this pre existing record
of saying please, and it knows that. And then it
knows I stopped when Sam said to stop. And I
think I'll get points for both of those decisions.
Speaker 3 (34:43):
See, I actually find if you're slightly meaner to the
models in your queries, they perform slightly better, much like
interacting with Joe.
Speaker 2 (34:53):
No comment on that. I you know what I do.
I have said luchadal like, I'll have a query and
at a respet bond and I'll say that was a
little on the nose, wasn't it? Like like it over.
Speaker 3 (35:05):
I'll say, do better.
Speaker 2 (35:06):
Yeah, I've seen that it does. Yeah, I've said like bad,
I've said things like that. It's like, this is really
on the nose. This response was a little bit trite,
don't you think, et cetera. Like I do feel like
I've gotten more comfortable at uh, let's be let's be
real here. You're not doing You're not doing your best
job here. What do you think is? Like, it's interesting
(35:28):
they like I get that they win the gold medals
and the math and et cetera. Like I've had so
many conversations with the chatbots that are on some level
like mind blowing what the capability is. I've never seen
a chatbot query that is like interesting, like that is like, oh,
that is like a really maybe one I could think of,
(35:49):
but there was like a really interesting thought. I feel
like my children still say on a daily basis more
like interesting things they get me thinking than I've ever
gotten from a chat Does that resonate to you at all?
Speaker 4 (36:02):
I don't know.
Speaker 2 (36:03):
I feel like you've said so many more things in
this hour than any than any interesting thing, like actually
like interesting ideas than I've ever gotten from the hours
I've spent playing with chat chipet your claude.
Speaker 4 (36:15):
You know, I use mine a lot for music. So
if I'm going to listen to Sibelius's Fifth Symphony, I'll
just ask it what should I listen for? And I'll say,
this is Tyler Kwan asking which I hope raises the
quality of the answer. It knows a lot about me. Yeah,
and what it gives me to listen for I find
is better than any human source I can access readily better.
Speaker 2 (36:36):
I that I agree with, but like, does it make
a connection, does it like tell you something about Sabilias's
music that is like, oh, that's a very interesting that's
a novel way of thinking about what makes it profound?
These are the things that I rarely ever encounter something
I was like, oh, that is an interesting thought, Whereas
I feel like if we're talking to a music cologist
(36:57):
for an hour, I would get infinitely more like actual
insight into something that makes the music special. So I
think I hadn't heard before.
Speaker 4 (37:05):
I don't know if it's novel because I don't really
know the Sibelius literature, but most musicologists I find pretty boring.
And I find say, GPT five on a classical symphony
quite to the point. And whether or not it's original,
it's not that important to me. It helps me listen
to the music better. Yeah, and it's certainly original relative
(37:25):
to the other sources at my disposal, say Wikipedia or
what I could google to So I think for almost
all purposes, that's enough. Does it have a truly original
idea in the sense that Einstein's theory of relativity when
he came up with it, was original to him? I
don't think so. That may come in some number of years,
(37:46):
but again, for almost all purposes, that's not what we need.
We need something better than our pre existing state of knowledge.
And on that I think it just cleans up.
Speaker 3 (37:55):
So I just ask perplexity what music I should recommend
to Tyler Collan, and it said that in order to
recommend effectively to Tyler Cowen, I need to look for
underappreciated recordings and obscure things that no one accept him
might have ever heard about. And it recommended. I mean,
(38:15):
boy genius seems pretty on the on the nose, right,
reggae acts like Toots and the May talent.
Speaker 1 (38:23):
This is so is that.
Speaker 2 (38:27):
Tyler Price records.
Speaker 3 (38:29):
So it's just scraping stuff that you've already talked to it.
Speaker 2 (38:32):
It's not even trying.
Speaker 3 (38:33):
Yeah, all right, well you.
Speaker 4 (38:35):
Need to make the prompt more exacting. Rule out anything
Tyler has talked or written about. Yeah, give me something
he doesn't know and try GPT five in the pro mode.
And I think it will succeed well.
Speaker 3 (38:47):
So on this note. This is something I've been asking everyone,
but like, what is an example in your mind of
a really good prompt or one that sticks out to
you that has generated something that you know, maybe you
didn't expect.
Speaker 4 (39:00):
You know, dwark Esh Patel once wrote a very good
prompt and he shared it with me. When I interview
some podcast guests, It's really a long prompt. It's in
hundreds of words, and it asks what questions should I
ask them? Then it goes into great detail. It should
be a unique question, it should be a question they
were not asked anywhere else. Then it's given what you
think their answer might be, and what would be my
(39:22):
follow up question? And yeah, give you give me some
cases where you think their answer might be wrong. And
it goes on and on, and you run that through
the very best models. I think you get good results. Yeah.
Speaker 2 (39:34):
No, I've you know, I've run transcripts of the podcast
before and I say, like, well, where should I have
pushed the guests harder on? What were the weak answers
that they what were the inconsistencies that the guests had
over the time? And I've found it to be a
very useful exercise for things like that. So I do
think like that's the thing which is first it is
(39:56):
it is objectively impressive on many of these from and
I would say objectively useful if you do sort of
a detailed prompting. I'm just curious, what do you what's
your like as a professor from the professor perspective, what
do you think is the right way to think about
how students will be using chadib I mean, I know
(40:17):
that there's a million opinions in academia about well, what's
the right way to test, now, what's the right way
to deal with essays, et cetera. How are you thinking
about some of these challenges.
Speaker 4 (40:28):
We should devote one third of all higher education to
teaching students how to use AI, and right now that's
close to zero, so we don't have the faculty who
can teach it. As part of the problem, often the
students no more than the professor.
Speaker 2 (40:43):
Yea, I'm sure.
Speaker 4 (40:45):
You need to restructure ratically what we do because future
work will be done with AIS, so that's the thing
to teach.
Speaker 2 (40:52):
But like so, just to play devil's advocate, like intuitively,
I still feel like there's value in long periods of
time cut off reading where you're not looking at devices,
where you're training your body to sort of be disciplined
and pay attention and focus. I still think memorization of
(41:13):
facts and numbers and dates and places and names is
very useful in actually having them in your head, et cetera.
Does that seem right to you or is that sort
of retro thinking on my part?
Speaker 3 (41:26):
No?
Speaker 4 (41:26):
Strong agree? And most of all writing, But that's the
other two thirds of higher ed right, I said, one third?
Speaker 2 (41:31):
Oh yeah, so tell us about the other two thirds.
Speaker 4 (41:34):
We should with or without AI just teach students much
more and much better how to write. Most people can't write.
Writing is thinking we should do much more to teach
writing and test writing, and now with AI, that has
to be face to face in a controlled environment where
people are just going to cheat, So that we should
have doubled down on to begin with.
Speaker 2 (41:55):
Yeah.
Speaker 4 (41:55):
So that and just numeracy and basic issues like how
to manage a portfolio, what kind of mortgage to take out.
There are classes that cover those things, but I think
they ought to be front and center of any curriculum.
Basic finance, basic life decisions like how to choose a doctor,
how to prompt the AI you know, for diagnosis whatever,
(42:16):
are relatively neglected in a lot of education. That to
me just seems crazy.
Speaker 2 (42:21):
I want to.
Speaker 3 (42:22):
Ask one market question before we go, which is obviously
there's a lot of talk about an AI bubble at
the moment, and I think the concern from a lot
of people is when you start talking about in new
technology as revolutionary, when you start talking about how you
know the effects are basically going to be infinite, and
the market size is hypothetically the entire world, there's a
(42:44):
risk that expectations overshoot reality, right and we have seen
some people voicing their worries about that right now, and
a little bit of nervousness creeping into the market. In
terms of valuations, Where do you stand on the AI bubble?
Do you see signs of frauth or do you think
most of the CAPEX spending is justified at this point?
Speaker 4 (43:04):
I don't like the word bubble. I would point out
that tech sector earnings are exceeding tech sector capital expenditure.
This is not mostly debt financed, so we're in less
trouble than many people think. It wouldn't shock me if
a lot of these efforts lost money. That was the
case with the railroads, the case with the internet, case
with most things humans have done. But I think it
(43:27):
will endure. It's not like pets dot com, where the
thing just gets swept away. These are incredibly well capitalized,
highly skilled companies where the CEOs and or founders are
quite committed to doing this and they're going to see
it through and they're going to succeed. But does that
mean every share value will go up or Nvidia ends
up being worth ten trillion dollars. I don't know. I
(43:49):
wouldn't necessarily predict that. There's always ups and downs, but
this is clearly a very useful thing and we're going
to we as Americans, we're going to make it work.
And we're way ahead of the rest of the world.
Like three quarters of all AI compute is in this country.
That's incredible. Worry what percent of the world's population six
or I don't know, but way smaller than three quarters.
Speaker 2 (44:11):
GPT five on Thinking Mode says you should listen to
Michael Gulisian, who does dream like acoustic guitar using an
open tuning, and it said that given your affinity for
guitarists like John Fahey and Leo Kottke, you'll appreciate him.
Speaker 4 (44:27):
Send me that answer. It sounds excellent. I haven't heard
of that person. Have you very much liked guitar with
open tuning, shure a.
Speaker 2 (44:34):
Monkar, a Hindustani vocal singer. It thinks you'll like sun
oh md Budrul Hawk. I'll send you this list and
Katerina Barbieri modern modular, minimist electronic composition.
Speaker 4 (44:48):
So I will buy some of these already.
Speaker 3 (44:51):
Wait, just how about a human recommendation? Oh, give us
a recommend a human recommendation. Since you said that you
know country is pretty good right now, but I guess
you personally aren't that into it. But have you tried
Orville Peck? I think there's a new album out out
this week.
Speaker 4 (45:06):
I think, what kind of music is it?
Speaker 3 (45:10):
Country? But like a very modern type of country. I've
tried to get Joe into it, but uh, I'm still
working on it.
Speaker 4 (45:19):
I like some country, and I love old country. So
Hank Williams, Johnny Cash, Jed Atkins period.
Speaker 3 (45:25):
It has a vintage toned to it, but with a
modern twist. Try Orville Peck and then get back to
us about which was better in terms of the recommendation.
Speaker 2 (45:36):
Tyler Colln Tyler Collen, thank you so much for coming
on odd Law's long overdue conversation. I really appreciate you
taking your time.
Speaker 4 (45:44):
Great to chat with you both.
Speaker 2 (45:58):
Tracy a lot to pull out from conversation. I think
it's very interesting that early observation he made about sort
of legacy institutions and whether perhaps some of the sort
of lack of revolutionary impact yet is just about that
metabolization process into the types of companies that could theoretically
(46:19):
absorb them.
Speaker 4 (46:20):
Yeah.
Speaker 3 (46:20):
I mean I think that's exactly it, right, So companies
are using this mostly as an add on to existing workflow.
You're not going to get the huge productivity boom until
companies are sort of centered from yeah, which you know
is probably going to take people who grew up with
the technology rather than old people like you and I.
Speaker 2 (46:40):
You and I, you just adopted it.
Speaker 3 (46:42):
The other thing I was thinking about, first of all,
insurers are sort of a pet interest of mine at
the moment, but I do think like they are emerging
as some of the really big winners from a lot
of like I guess, the data saturation of the world
right now and the increased sophistication of analytical models and
things like that, and that'll be very interesting to see
(47:02):
how it shakes out. And you could if you if
you took it very, very far as a sort of
thought experiment, you could start to say that, like, well,
the insurers are going to be a more important actor
in terms of setting social standards and regulations in the future,
because they're the ones with the data, doing all the
modeling and saying like, you have exactly an ex chance
(47:23):
of being in a car accident, and therefore you must
do the following things, right.
Speaker 2 (47:28):
I thought also, like I hadn't really thought about, you know,
subpoena ability. I think it is a very big issue,
but it is interesting, right, Like, it is a little
weird that you and I could have a phone conversation
now if you're under oath and they say, what did
you talk about? Joe? I expected you'd probably tell the
truth unless it's very bad for me, and I would
(47:50):
hope that you would lie.
Speaker 3 (47:52):
That's right.
Speaker 2 (47:52):
Yeah, I hope that you would. I would hope that
you that's your hope. I would hope that you myself
to save Joe. Yeah, I would hope that you would
perjure yourself, but like you know, theoretically could get away
with it. So yeah, well we have an email, there's
no chance. And it's sort of interesting. It's sort of
it's obviously it seems a little bit arbitraryed me, but
it is interesting to think about, like, Okay, can we
have a conversation with these entities? And like why do
(48:15):
we have to leave a digital record? And where are
these going to be stored? And I do think in
areas like health and law, which are obviously not obviously
but sort of intuitively low hanging fruit for productivity gains,
how much have we not seen, just in part because
we're still sort of negotiating the transition process as the society.
(48:36):
What are going to be the new rules and norms
about this stuff, where it's gonna be hows, et cetera.
I think that's actually a sort of very interesting question.
Or space space, This is the.
Speaker 3 (48:45):
Space to watch, yeah speak, Now that I think about it,
we probably should have discussed some of the regulatory framework
around all of this a little bit more, But next time,
next time. But the other thing I've been thinking about
lately is economic statistics. Yeah, in a world that's increasingly
driven by AI, and I know that we had the
big productivity discussion and technology in relation to technology in
(49:08):
like the sort of early to mid two thousands. I'm
not sure that ever actually got settled, but I very
much expect that the AI economic statistics conversation is going
to be even like wackier because I'm not sure how
you do things like quality adjustments for something that like
suddenly comes with its own brain and stuff like that.
Speaker 2 (49:29):
So no, it's gonna be super weird. I did, I
don't know. In my mind, I have like a very
these images of like Tyler getting a tour through the
you know, chatch EPTE offices and the people ask you
was like, well, and him having to explain that, you know,
we're not going to have twenty percent GDP growth at
maybe two and a half or sorry productivity growth, and
(49:50):
actually GDP isn't really that bad of a measured more
or less captures the size of the economy, even if
a lot of Internet things are free. Like some of
these classic convers stations, I would like to be a
fly on the wall for some of those.
Speaker 3 (50:05):
Tyler Cowen in defense of GDP dependent Yeah, all right,
shall we leave it there.
Speaker 2 (50:10):
Let's leave it there.
Speaker 3 (50:11):
This has been another episode of the Odd Lots podcast.
I'm Tracy Alloway. You can follow me at Tracy Alloway.
Speaker 2 (50:16):
And I'm Joe Wisenthal. You can follow me at The Stalwart.
Follow our guest Tyler Cowen, He's at Tyler Collen and
of course check out his podcast Conversations with Tyler an addition,
of course, Marginal Revolution. Follow our producers Carmen Rodriguez at
Carman armand dash O Bennett at Dashbot and Cal Brooks
at Cale Brooks. From our Odd Lots content, go to
Bloomberg dot com slash odd Lots with the daily newsletter
(50:39):
and all of our episodes, and you can chat about
all of these topics twenty four seven in our discord
discord gg slash odlines.
Speaker 3 (50:47):
And if you enjoy odd Lots, if you like it
when we talk to Tyler Cowen and give him music recommendations,
then please leave us a positive review on your favorite
podcast platform. And remember, if you are a Bloomberg subscriber,
you can listen to all of our episodes absolutely ad free.
All you need to do is find the Bloomberg channel
on Apple Podcasts and follow the instructions there. Thanks for
listening
Bloomberg Audio Studios, Podcasts, radio News.
Speaker 2 (00:18):
Hello and welcome to another episode of the Odd Lots podcast.
Speaker 3 (00:21):
I'm Jolle Wisenthal and I'm Tracy Alloway.
Speaker 2 (00:25):
You know, not a novel observation here, but I do
think that if, like in early twenty twenty two or
late twenty twenty one, you had someone had revealed to
you all of the amazing things that we could do
with AI these days, I think you would have expected
that either the broader economy or society would be more
different than it has been. Like I am large, you know,
(00:49):
I think most people's jobs are done roughly the same way.
Sort of society still seems to operate, although a little
maybe a little bit worse every day. I don't know,
But like I just don't think like it's mind blowing
technology and get by and large, like it hasn't had
the economic disruption that I think many people would have guessed.
Speaker 3 (01:07):
Maybe I'm cynical on the subject, but I always say,
never underestimate the human capacity for stasis, I guess, and
making things much more difficult than they actually need to be,
and putting up you know, bureaucratic barriers, regulatory barriers and
things like that. So it's not it's not that surprising
to me, but it is true. You have a lot
of economists out there who think there was going to
(01:27):
be this massive productivity boost, right well, it's certainly.
Speaker 2 (01:30):
Tech people like I don't know, like economists. You know,
they're like, oh, there were some we're going to have
a productivity boom. We're raising our estimates from two percent
too and a quarter percent, and then you have all
things relative to and then you know, we talked to
Kathy Wood one time, and I think she what did
she predict twenty percent real GDP growth for twenty years?
Or you certainly have these people in Silicon Valley deflationary boom,
(01:54):
post scarcity any any minute, right now, some real gaps
I would say, between how lot of economists think about
some of these things and the numbers, then economists would
be comfortable using versus maybe literally.
Speaker 3 (02:07):
Everyone else like Kathy would certainly many others. Yeah, you
know what I think about is how different would your
career in my career have been had AI existed in
like two thousand and eight, two thousand and six when
we were starting to blog.
Speaker 2 (02:22):
Basically, it's a really good question, and I don't know
the answer because part of me thinks, well, you know,
that whole path that was like defined my career like
would not have been there, would not have existed or no. Maybe,
But on the other hand, maybe in two thousand and
six I would have been like just like I was
super early into blogging, super early into experimenting with AI news,
(02:45):
and it would look different, but maybe I would have
rowed some different.
Speaker 3 (02:48):
Different but the same.
Speaker 2 (02:49):
Yeah, different but the same, you know, Like I think
it's a it's pretty hard to tell.
Speaker 3 (02:53):
One thing I know is like back then we were
writing to optimize Google results, right, that was basic or know,
like I know social media was disseminating stuff as well,
but like Google was still the main platform for a
lot of stuff. I'm sure if we were doing it now,
we would be optimizing for chat, GPT or perplexity or
someone like that to actually pick up the content. I mean,
(03:16):
I think that's the audience.
Speaker 2 (03:17):
Yeah it is. And I think like any especially any
like commercial publisher, right that's trying to do a scale.
I think like a sort of niche voice expert could
still have like their audience that comes to them directly,
of course, but like it's scale scale for sure every
publisher is trying to figure this out. Anyway, We really
do have the perfect guest, someone we should have had
on the podcast years and years ago. It's almost surprising
(03:39):
that it's just the first time we've had him on.
It is crazy, It is crazy, but it is someone
who I think is like at the intersection of everything
that we're talking about. Because he's an economist. He knows
the tech really well. The tech people know him, the
tech people love him. He may even be one of
the preferred economists for this. I feel he's like the
economist that all people in the AI world want to
(04:01):
get his take. A long time blogger, someone I, you know,
like both of us sort of gave up on mike blogging.
Although we have our newsletter. It's like close enough, but
it's different than blogging. Someone who's stuck with the medium
for a long time. One of the true original econ
bloggers that I've been reading for over twenty years. We're
going to be speaking with Tyler Cowen. Here's the host
of the Conversations with Tyler podcast. He also, obviously is
(04:22):
one of the two bloggers that the famous Marginal Revolution blog.
Economics professor gmu appreciator of ethnic foods all around the
DC Northern Virginia, Maryland area. Someone known on the internet
for a long time, Tyler, thank you so much for
coming on odd Laws. Really thrilled to have you here.
Speaker 4 (04:44):
Hello, happy to be here.
Speaker 2 (04:45):
Amazing. What do you think about my initial assessment? How
fair do you think it is that, Like, had you
known in twenty twenty one how powerful these tools would be,
maybe we'd be a bit surprised that by and large,
like busines, and this seems to more or less run
the same, I'm.
Speaker 4 (05:03):
Not surprised at all. So what I see right now
is people using AI as an add on to their
pre existing work routines. Oh, you need to write a memo,
you ask AI how to do it. You need to
write a column, you ask AI to proofread it or
fact check it. And that works great, But those are
marginal gains. What we really need to see a major
impact is new organizations built around AI, and those will
(05:27):
be startups. They will come only slowly. It will take
twenty or more years before they really transform the economy.
And in the meantime, it's a whole bunch of add ons,
which are fun and fine, but that's why I think
it's slow.
Speaker 2 (05:40):
What can you tell us about the history of technology
such that legacy organizations that existed prior to the invention
of maybe potentially revolutionary new technology have a hard time
massively changing their workflows.
Speaker 4 (05:58):
Well, you can take even very simple examples. So Toyota
starts competing with General Motors in the nineteen seventies, General
Motors is paralyzed. It cannot really come back and adopt
the new and superior Toyota methods, and those are not
really that different compared to say AI. So there's just plenty,
plenty examples. Old mainstream media could not cope very well
(06:20):
with the Internet. There are exceptions like The New York Times.
Odd Lot's podcast would be another thank you, But it's
the norm. So we're seeing it again with AI. And
again you need a complete turnover of who and what
is doing business for it to really matter on a
big scale.
Speaker 3 (06:37):
If you need that complete turnover, and you need some
time for AI to become fully embedded in a business
model or for a business model to form around AI,
what industry or what part of the economy would you
expect it, I guess to show up first in that
sort of revolutionary way.
Speaker 4 (06:56):
Well, we have obvious data on this, and it's programming.
You will hear people who do programming claim that say
eighty percent of the work is now done by AIS.
I suspect that's an overstatement, but there's no doubt at
all that there's simply a lot of programming already done
by AIS. When you have low fixed costs, a competitive sector,
immediate feedback, you know, the revenue has to flow programming.
(07:21):
And also New York City finance, there's been quantston finance
for a long time. Those quants are now you could say,
more AI equipped than they used to be, and those
areas already are being revolutionized.
Speaker 2 (07:33):
I've heard them and got some pitches to do episodes,
which we should do at some point. But I've heard
about like some law firms that are being like new
AI law firms where there are lawyers, et cetera. But
from the very ground up, the idea is perhaps there
is some way to just get way more productivity if
(07:54):
they start from the very beginning with some combination of
lawyers plus AI models. It seems like that could be
the kind of thing where maybe the legacy law firms
are seeing some productivity gains from AI. There's probably some
evidence you could find but that a new one with
a totally different approach could deliver that big productivity boost
(08:14):
that actually ends up changing the industry.
Speaker 4 (08:17):
It's already the case, say, mid tier associates are much
less needed. But there's one big problem with law in particular,
and that is the way large language models work. Now
you have to send your queries somewhere else. You can't
just own and control and hold the whole thing on
your hard drive. Now. I think within a few years
time that will be very different. But until then, major
(08:38):
law firms are extremely skittish about just typing in their
questions and sending it, you know, to San Francisco. I
don't actually think there's a risk, but when you think
about how fiduciary responsibility works, they just don't want to
do it.
Speaker 3 (09:07):
Talk more about I guess privacy concerns and regulation, because
this seems to be an area where if you are
in a heavily regulated industry, or if you're just in
an industry that tends to be full of paranoid people
like lawyers, it does seem like there's going to be
a natural tendency to be very very cautious. When it
comes to sharing data with AI, you're going to be
(09:28):
worried about actual data ownership the queries that you're sending
to San Francisco, as you say, are those industries just
inevitably going to be slow to adapt.
Speaker 4 (09:37):
They'll be slow to adapt again until the point where
they just control their own model and they hold it
within the firm and no one else really can get
at it. So what you need is cheaper models where
a law firm can afford to have its own model.
And I think that's a few years away. It's not
a very long time away, but it won't come in
(09:59):
six months. Sam Altman I just did a podcast with
him and he said a privacy problem is AI queries
or subject to subpoena, and he thinks they should have
as much protection as say, your conversation with your lawyer
or your doctor or your therapist. I think that's a
good idea, but that hasn't happened yet, and until it happens,
(10:19):
or you get the whole thing on your own hard drive,
progress in law is going to be slow. But once
the progress comes, that's one of the areas where I
think AI has the most promise it's just very very
good at mastering a large corpus of text and organizing
it for you.
Speaker 2 (10:35):
It is interesting, isn't it. Where it's like, Okay, if
you are my lawyer, you and I could have a
conversation and it would be not it would not be uh.
Speaker 3 (10:44):
It'd be privileged.
Speaker 2 (10:45):
Yeah, it would be privileged. Or you know, one can
have conversations with anyone, and as long as you're doing
it on the phone, or if you're doing it person
to person, it's much more difficult to get that, you know,
in discover or I always think about this with you know,
public public records of employees that sometimes you can get
(11:06):
their emails through Floyers, et cetera, but you can't get
the cons content of conversations. It does feel I hadn't
really thought about this dynamic though, that when you're using
an AI it is sort of like a conversation and
yet sort of from an evidentiary basis, it would be
much more like an email.
Speaker 4 (11:25):
I think when it comes to medical issues, there are
many more people willing to share their data. Not everyone
by any means. Some medical conditions are secretor people just
don't want others to know. But I see many, many
people I know typing in all kinds of things about
their medical history to say GPT five and getting what
are on the whole very good answers. It's like medical
(11:45):
diagnosis for free spreading now to the whole world. A
lot of countries where people just don't even have access
to good doctors at all, and I think that will
be important more quickly than the law innovations.
Speaker 3 (11:59):
Just as a thought experiment, what does all of this
mean for insurers? Because I kind of think, you know,
I think about a bunch of people typing in their
medical information. I think about basically the explosion in data
that we have nowadays, and it seems like the insurance
industry would be one place that would really benefit from
all this trove of additional information if they could access
it well.
Speaker 4 (12:21):
This is one of my worries about AI in general.
I'm quite positive on what's happening, But as insurers get
better and better information on their customers, this is just
through big data more generally, it doesn't have to be
current large language models. They know exactly how to write
the risk and how to price the premium, and in
a sense for the buyer, it's not insurance anymore. So,
(12:41):
if we know your house is going to burn down
with high probability and you have to pay the super
high premium, you don't really have the benefits of the insurance.
So some insurance markets might unravel if through big data
the insurers learn too much about what's likely to happen.
Speaker 2 (12:57):
Economists seem to be very consistent about the effects of
technology on labor demand, which is that in the end
it washes out right, some people, Okay, there's a disruption,
but I'm gonna save money because I use the I.
But that means I have more spending power, and then
I'm gonna buy something else that I wouldn't have bought
had it not been for paying wages, and then that'll
(13:20):
create demand for labor elsewhere. And so in the end,
the idea that you could really have tech driven unemployment
at scale that is not transitory or not temporary is hard.
A lot of economists seem to be intuitively skeptical of
this idea. Whereas you have people in the ifields, fifty
percent of the people aren't gonna have jobs we need
(13:41):
ubi otherwise there's gonna be a permanent underclass. Could there
be something different about AI such that it doesn't have
the same labor market effects that past technologies have had.
Speaker 4 (13:52):
I would say, you understand me. Well, So the energy sector,
there's going to be a lot of new jobs taking
care of older people. I think as AI produces more potential,
medical innovations will need to test them, So the biomedical
sector testing clinical trials, there'll be a lot new jobs.
I'm not worried about mass unemployment, and most economists are not,
(14:14):
and I agree with their perspectives, which I think you
outlined pretty clearly.
Speaker 2 (14:18):
What do you say, though, because I have a feeling
that when you're out in San Francisco, they don't see
it that way, and they talk a lot, and some
of them are more in you know, UBI talk and
permanent underclass talk and all of this stuff. Do they
see your perspective when you make this case? What do
they say or what are they What are a lot
of them seem to be missing about the logic that
(14:39):
you spell out.
Speaker 4 (14:40):
Well, I think more and more they're coming around to
the economist's point of view. So Andre Carpathi, who was
at open AI, you know, in the most important years,
he just did a podcast saying he thinks adjustment will
be slow. Things will be fine, will grow at two
points something percent. There won't be mass unemployment, and you
wouldn't I have heard that, say two years ago. But
(15:02):
I think as people see the models rolling out, and
as you mentioned well, the real world impact it is
stretched out in time, right, it's not all immediate. Earlier
on people had more the sense that AI was a
kind of god box that you just talked to it
and it can magically do anything and convert that into
results in the real world. But if you think about
(15:22):
your actual job, even if it's a highly intellectual job,
so much of what you do is the interaction between
your intellect, your physical presence, your interactions with others, your travel,
many other things. And until we get to some far,
far off world where the robots are perfect copies of you,
which I don't think ever will come, jobs will be fine,
(15:43):
but they will change a lot, and I'm actually worried
about who will be the biggest losers. I think poor
people will do great, the very wealthy will do well,
but people who are sort of upper upper middle class
will find this automatic ticket to a law or consulting
job that assured they would be upper upper middle class
(16:03):
for the rest of their lives. I think a lot
of that is going away already.
Speaker 3 (16:08):
I think also no one would have expected plastic surgery,
I guess, to be a beneficiary of the AI revolution.
But if you think that what's going to matter in
the future is like your personal presence and your network
of social contacts, then I guess we should all be
working on on our looks Maxing, Yeah, absolutely right, charisma. Okay, noted,
(16:29):
everyone work on their respective charisma. One thing I was
wondering is the impact of AI on public finances. I
am not very good at tax policy. Joe knows this
because I've complained about taxes to him repeatedly. But I'm thinking,
but if you're thinking about where the value add of
(16:49):
AI actually shows up in the economy, So you know,
presumably you got a productivity boost. We're not entirely sure
how much that's going to be. But where does that
additional output actually show up in terms of revenue for governments?
How is that collected? And how would you expect the
distribution to vary across the world.
Speaker 4 (17:07):
Well, I think in the United States medium term, they'll
just be much much more. Healthcare will be new drugs,
new medical devices, will have to test all these things,
will have to produce them, and that will be It
was already the case that those sectors were growing, but
that growth will be accelerated. So that's where I think, God,
they'll be the biggest difference longer term, and people will
(17:29):
live longer because we'll fix at least partially various diseases
and maladies. So if you live to be ninety four
across a lifetime, you spend way more on healthcare than
if you live to be seventy seven. And that's yet
further growth for the healthcare sector. But some things like
medical diagnosis that's already very cheap, like you know, a
(17:50):
good large language model probably outperforms your current doctor at
least if you type in what's wrong with you properly.
Speaker 3 (17:57):
But does that additional productivity or or the output generated
by AI, does that actually show up an additional taxation
for the government.
Speaker 4 (18:06):
Well, the healthcare sector generates an enormous amount of taxation revenue.
I do think we'll have some sectors that maybe just
become free in the same way that Wikipedia is free.
So I could imagine say ten or twenty percent of
the music sector is music you create at home using
your own AI and it's a customized song for you,
(18:26):
and maybe you paid a subscription for the service. But
rather than spending more money on Spotify or a streaming service,
you just build the music and that's a partial substitute
for some human created music. I don't think human created
music will go away at all. People want to enjoy
the human touch, the feeling that you're a fan of Taller,
Swift or whatever. But there's going to be a lot
(18:48):
of AI generated music and art and many other areas,
and some of it will be free. But that's not
a problem from a revenue standpoint. So instead of spending
money on safe buying a picture, you create one digitally
at home with your AI. You'll spend that money on
something else.
Speaker 3 (19:06):
I got to ask, now, are you a swiftie.
Speaker 4 (19:09):
No, it's worrying for me. It's a little too predictable,
so I have to say I'm not, but I'm glad
other people are. Let's put it that way.
Speaker 2 (19:19):
Do you have a theory let's talk, let's just talk
about music. Do you have like a theory of the
swift phenomenon? Like, is there a reason? What is it?
Because it's just so, doesn't I don't know what's your
what's your meta take on the tailor, the swifties and
culture and society.
Speaker 4 (19:38):
Well, it's super polished and because of the Internet, the
very biggest of celebrities can be much bigger than before,
and someone will fill a few of those slots. But
I think also how she presents herself. She has the
guise of being attractive without feeling threatening to other women. Uh,
and there's something all American about her and quite generic.
(20:00):
She doesn't rule out the fandom of that many people
m hm. And she's the one who's filled that slot.
She's been brilliant at managing her career and seems to
just stick at it and has an incredible work ethic.
People say the shows are incredible. As more and more
life goes online, who can give a good show? It's
the charisma and looks. Point Well, she seems a pleasant
(20:21):
that I've never been to one, but I hear plenty
of reports. And you put all that together and she's,
you know, the megastar of the music world.
Speaker 2 (20:29):
Do you think like culture, Like there's this popular idea
that culture is sort of dead, and I do think
that is probably overstated, But you know, you look at
movies and people have people have observed this for a
long time. It's just rehashes of franchises that have been
around for thirty years. And I think if you look
at Spotify's streaming, there's still this overwhelming tyranny of the
(20:52):
boomer rock et cetera, and this feeling that culture in
many respects is that's rehash that is very hard for
new things to to break out. I mean, Taylor Swift
at this point is decades old phenomenon. Is that real
in your view? Or is that just people in the
pundit class who've been lazy and not discovering new things
(21:13):
and not actually putting in the effort because they're not
young anymore and they're not going out and they say
nobody listens to new music.
Speaker 3 (21:19):
Joe, are you just talking about yourself?
Speaker 2 (21:20):
I have talked about I'm trying. I'm doing a little
introspection here. Is this just me because I don't go
to shows like I did when I was twenty or
is something changed?
Speaker 4 (21:28):
But a lot of it is the pundit class. So
you look at movies, I think it's perfectly correct to
say the most popular movies today are pretty dreadful, and
it used to be the case that the most popular
movies were God, The Godfather in Star Wars. Yeah, that's
a big change. But if you look at movie making
around the world and in a given year list say
the twenty five best movies from all places, which may
(21:51):
not even come to your multiplex every year, you have
an incredible list. I don't think they're worse than the
movies of earlier times. I do think mainstream hollywo is
much worse. So in many areas you just have quality
moving more into unises.
Speaker 2 (22:06):
By the way, I just want to say, the first
time I ever encountered your work, prior to even having
stumbled across Marginal Revolution, was at the bookstore in Austin
finding a copy of In Praise of Commercial Culture, and
I just feel like that book is held up so well.
I mean in the specific sense that there is so
(22:26):
much mass culture these days, whether it's high end Netflix
TV shows, et cetera, whether it's A twenty four films,
whether it's Beyonce or Taylor Swift or some of these
other big names who are simultaneously incredibly popular. And I
get that you're not as swifty, neither mine anymore, but
people take this yeah, I liked their country for once
(22:49):
she left country. But like where people take these popular
culture things extremely seriously as art and don't dismiss these
outputs as sort of being trash.
Speaker 4 (23:02):
Right now, we're in a golden age for country and
western music and also horror movies. Neither is really like
my taste in particular, but it's easy to see what
has gotten worse, and especially for critics, harder to see
what has been getting better.
Speaker 3 (23:17):
That's my sweet spot. Country and horror is perfect for me.
But just on the culture point, I mean, I think
the lack of culture argument, the one I hear the
most is it's a lack of shared culture.
Speaker 4 (23:28):
Right.
Speaker 3 (23:28):
So you do have like some giant monolists like a
Swift or a Beyonce that everyone knows about, and you know,
they do have these large audiences, but broadly we're not
all experiencing the same media that we used to, right, Like,
no one is gathering around the TV to watch the
finale of you know, some show that airs like once
a week and has been going on for five years.
Speaker 2 (23:49):
The exception is NFL football, which I don't really get it,
but yeah, but it's not cultural.
Speaker 3 (23:53):
Sports is outside of my experience.
Speaker 4 (23:56):
And the Super Bowl is cultural, right, it's a cultural event.
You care about the advertisements, But the biggest of YouTube stars,
which again I would say is not personally my thing,
but they can have bigger audiences than those older TV shows. Right.
Speaker 3 (24:10):
Well, So what I'm getting at is it does feel
like nowadays there's an ability because of tech, to serve
up very specific content and like niche content in streams.
And the analogy that I like to use is, you know,
everyone knows if you download Netflix for the first time,
the first movie you watch is incredibly important because whatever
(24:33):
you watch, you know, if it's a rom com, you're
going to be served up Kate Hudson films for the
rest of your life. Right, Like, the algorithm looks at
what you're watching and then it serves up that additional content.
What does that mean for society? The idea that you
have people you know, basically funneled into smaller and smaller
streams in some respect, well.
Speaker 4 (24:54):
A lot of the Netflix algorithm it just directs you
to slop. True, people have always wanted slop. Like people
listen to music way back when it was quite common,
or they would just list in the top forty, which
in some years was very good, but often was pretty terrible,
even in the nineteen sixties. So what you can do
today is basically watch not any movie out there, but
(25:18):
you have movie You can still buy DVDs and Blu rays.
You have access to more cinema today than you ever have.
So people will sort themselves. And I think it's from
the point of view of cultural consumption. I don't think
there's ever been a better time to be alive than
right now. Like, well, a lot of people abuse that
(25:39):
and go for slop. Of course that's sad, but it's
hardly new.
Speaker 2 (25:43):
So changing gears a little bit. Tracy and I write
almost every day because we have a daily newsletter that
forces us to and I really like having that because
I don't know if I would write every day if
I did not have that obligation to deliver something in
people's inbox that they pay for as partler Bloomberg dot
Com subscription. I love writing, but I don't know if
I would do it every day if I didn't have
(26:05):
this sort of requirement. I might just tweet. How do
you You've been blogging for over twenty years, how do
you avoid the temptation to just fire off all your
ideas via tweet and actually commit to the blog.
Speaker 4 (26:19):
I'm never tempted to do that. I like to think
things out.
Speaker 2 (26:22):
I don't do.
Speaker 4 (26:23):
I just don't more when I write properly. I've actually
blogged every single day for over twenty two years.
Speaker 2 (26:29):
That's amazing. It's amazing. Most people gave up, and so
what's the difference.
Speaker 4 (26:34):
I don't feel it requires any discipline. For me. The
discipline is not writing more like I have to restrain myself.
So I guess I'm just weird. I don't think I
have any neat little trick or formula. It's one of
these niches that you can do now that you couldn't
do before. And I found my niche, as have the
two of you.
Speaker 3 (26:53):
Here's a slightly different question playing on that theme, but
going back to the intro, How different do you think
you're blogging career would have been had chat bots existed,
you know, ten or twenty years ago when you were
starting out.
Speaker 4 (27:07):
I don't think we know yet. My intuition is that
people still want to read human writers simply because they're human,
and if the bot is as good as you, most
of the world doesn't care. But that has not truly
been tested yet. I think we'll see in the next
two years, but that's what I'm expecting. Just I think.
I think in music, there'll be plenty of AI music.
(27:29):
It might be say ten or twenty percent of the
music sector, but listeners will still want that human to
human connection.
Speaker 2 (27:51):
Do you think you know, when Twitter came out, they
called it a micro blogging site, as if it were
just blogging but on a shorter basis. But I think
it's fundamentally different. You know, in the early, the glory
days of blogs, which we'll be talking about forever when
we're all very old people, how good it was, you know.
I think there was this sort of spirit of you know,
(28:11):
liberal linking with each other and idea exploration, whereas Twitter
strikes me as much more conflictual and one upmanship and
so forth. Do you think there are like fundamental I
don't know if political is the right word, but like
new communication paradigms like sort of have their own terroir,
so to speak, in terms of the impulse towards collaboration
(28:34):
or conflict, etc. And does that change society?
Speaker 4 (28:38):
Yeah, I still like blogging, and I'm sad people have
moved away from it Twitter. To me, it seems too
meme heavy, and meme heavy media have more potential for racism,
which of course is a big negative. And I see
so many people who are driven crazy by being on Twitter.
Whether it's because they're writing on it or reading it,
I'm not sure, maybe both. I don't want to name names,
(28:59):
but it's a lot of people, and I bet you
see the same ones that I do.
Speaker 3 (29:03):
Also very sexist nowadays, I would just add in unappreciated
ways in many ways speaking of sexism, the impact of
AI on economics. Talk about that economic institutions, you know,
famous for modeling, spend a lot of time with numbers
and things like that. Is that all just going to
be replaced by AI?
Speaker 4 (29:23):
Not all? So. I think what human economists will do
is put more and more time into gathering data and
feeding it to the AIS. The returns to doing that
will be very high. But the actual econometrics, statistics humans
will maybe set up part of the problem, but the hard, boring,
routine work will be done by the machines. As to
(29:43):
some extent it was already the case, and this will
be a way to make a lot of progress rather
quickly do the.
Speaker 3 (29:49):
Actual economic statistics or data points that economists collect. Do
some of those need to be changed or thought of
differently in light of the AI era.
Speaker 4 (30:00):
Well, I don't know. Eventually they will need to be,
but I would say any period of radical change in history,
your statistics are less useful. It's not that the people
creating the statistics are making some mistake. You just cannot
capture every way the world is changing, and index number
comparisons require the basket of goods be relatively close to constant,
(30:23):
and at some point that doesn't hold anymore, and we'll
be faced with that. We'll deal with it. I would
say the current statistics we have, they're more underrated than overrated,
so they're actually pretty good. I'll be glad when we
get them back again.
Speaker 2 (30:38):
There's another thing. The tech people you talk to, they
must think like GDP is terrible. Doesn't capture all this stuff,
all this value that we can't price. I'm sure you've
had conversations explaining to many AI workers that GDP is
not the worst statistic in the world. I want to
go back though, to them. So you know you mentioned
like on Twitter, right, you say something someone so it's
(31:00):
a meme. They dunk on you, they make fun of you,
they whatever. Your chatbot won't do that. Like if I'm
having a conversation with Chad GBT, it's never going to
respond to me with the meme sort of indicating one
day of my ad that I'm.
Speaker 3 (31:14):
A moral you'll learn your language, Joe.
Speaker 2 (31:16):
Yeah, But if anything, it is too obsequious, right, I mean,
the issue is like the online world has become this
very like sort of competitive, conflictual world, and then I
go to the chatbot and my complaint is literally the opposite.
It doesn't challenge me enough. It's too obsequious. Every question
I ask, it's a great question. Sometimes I wish it
would call me a moron a little bit more. But
what does it change about the world? And we know
(31:38):
that like many people's brains have been broken by social
media that probably has downstream effects on how our politics
operates these days. What does it do to the world
if we started inhabiting these chat environments where they're just
very sort of polite, and every time you say something,
it says, yes, great thought, Tyler, great thought, Joe, would
you like to expand it? You asked the perfect question,
(32:00):
what do you see that having sort of second order
effects on how society operates.
Speaker 4 (32:05):
Well, that's the four to zero model that does that.
The newer models like Claude four point five and GDP five,
they're more objective, and that's better.
Speaker 2 (32:13):
But they never make fun of you. They never will
like say, you're a moron? How could you possibly how
could you possibly ask such a dumb question? You're obviously
so out of touch for having the ax ask this.
This is like a very in some respects. This is
a very positive change from many of the conversations that
I've had from typing into a computer.
Speaker 4 (32:34):
Oh it's great. I think people should be nicer to
each other. And I think they're the most objective media
source the human race ever has had. If you're ask
good about, say vaccines or conspiracy theories, it basically gives
you the right answers.
Speaker 3 (32:48):
Well, one thing that they don't do, and I mean
I do think they can be trained to be mean
to you and to insult you to a certain degree,
and you can.
Speaker 2 (32:56):
Could bring but I never account of them.
Speaker 4 (32:58):
Try market that more readily on PAP right.
Speaker 3 (33:02):
So we spoke to the chief business officer of Perplexity,
Dmitri Shevalenko. We spoke to him recently, and he was
saying that one thing chat models can't do is express
a natural curiosity, which I thought was kind of weird
coming from him, because perplexity is the only model I
know that actually throws out those additional questions if you
query it and then it comes up with would you
(33:23):
like to have more information on this point or are
you thinking about this now? But there does seem to
be an element of creativity perhaps that is lost in
some of these lms. How much does that change things
in media? The idea that you know, the models are
going to spit out something that's sort of predestined in
many ways.
Speaker 4 (33:44):
Well, that's what most humans do, to be clear, But
it's now the case that on a regular basis, the
models say can prove new theorems and math or discover
new potential drugs. And keep in mind, you know, a
year ago these things thought the word strawberry had too
and now they're winning gold medals and math Olympias. So
a year or two from now, maybe we don't know
(34:06):
how much better they'll be, but I don't think they're
gonna have any problems being creative, certainly more creative than
humans on average.
Speaker 3 (34:14):
So now I have to ask. Since we're talking about
being mean or nice to the models and them being
mean or nice to you, do you say please and
thank you in your LLM queries?
Speaker 4 (34:23):
You know? I used to, and then Sam Altman said, well,
it costs us just a little bit of money because
of the extra tokens. And then I thought, I'll hold
off on this. But I have this pre existing record
of saying please, and it knows that. And then it
knows I stopped when Sam said to stop. And I
think I'll get points for both of those decisions.
Speaker 3 (34:43):
See, I actually find if you're slightly meaner to the
models in your queries, they perform slightly better, much like
interacting with Joe.
Speaker 2 (34:53):
No comment on that. I you know what I do.
I have said luchadal like, I'll have a query and
at a respet bond and I'll say that was a
little on the nose, wasn't it? Like like it over.
Speaker 3 (35:05):
I'll say, do better.
Speaker 2 (35:06):
Yeah, I've seen that it does. Yeah, I've said like bad,
I've said things like that. It's like, this is really
on the nose. This response was a little bit trite,
don't you think, et cetera. Like I do feel like
I've gotten more comfortable at uh, let's be let's be
real here. You're not doing You're not doing your best
job here. What do you think is? Like, it's interesting
(35:28):
they like I get that they win the gold medals
and the math and et cetera. Like I've had so
many conversations with the chatbots that are on some level
like mind blowing what the capability is. I've never seen
a chatbot query that is like interesting, like that is like, oh,
that is like a really maybe one I could think of,
(35:49):
but there was like a really interesting thought. I feel
like my children still say on a daily basis more
like interesting things they get me thinking than I've ever
gotten from a chat Does that resonate to you at all?
Speaker 4 (36:02):
I don't know.
Speaker 2 (36:03):
I feel like you've said so many more things in
this hour than any than any interesting thing, like actually
like interesting ideas than I've ever gotten from the hours
I've spent playing with chat chipet your claude.
Speaker 4 (36:15):
You know, I use mine a lot for music. So
if I'm going to listen to Sibelius's Fifth Symphony, I'll
just ask it what should I listen for? And I'll say,
this is Tyler Kwan asking which I hope raises the
quality of the answer. It knows a lot about me. Yeah,
and what it gives me to listen for I find
is better than any human source I can access readily better.
Speaker 2 (36:36):
I that I agree with, but like, does it make
a connection, does it like tell you something about Sabilias's
music that is like, oh, that's a very interesting that's
a novel way of thinking about what makes it profound?
These are the things that I rarely ever encounter something
I was like, oh, that is an interesting thought, Whereas
I feel like if we're talking to a music cologist
(36:57):
for an hour, I would get infinitely more like actual
insight into something that makes the music special. So I
think I hadn't heard before.
Speaker 4 (37:05):
I don't know if it's novel because I don't really
know the Sibelius literature, but most musicologists I find pretty boring.
And I find say, GPT five on a classical symphony
quite to the point. And whether or not it's original,
it's not that important to me. It helps me listen
to the music better. Yeah, and it's certainly original relative
(37:25):
to the other sources at my disposal, say Wikipedia or
what I could google to So I think for almost
all purposes, that's enough. Does it have a truly original
idea in the sense that Einstein's theory of relativity when
he came up with it, was original to him? I
don't think so. That may come in some number of years,
(37:46):
but again, for almost all purposes, that's not what we need.
We need something better than our pre existing state of knowledge.
And on that I think it just cleans up.
Speaker 3 (37:55):
So I just ask perplexity what music I should recommend
to Tyler Collan, and it said that in order to
recommend effectively to Tyler Cowen, I need to look for
underappreciated recordings and obscure things that no one accept him
might have ever heard about. And it recommended. I mean,
(38:15):
boy genius seems pretty on the on the nose, right,
reggae acts like Toots and the May talent.
Speaker 1 (38:23):
This is so is that.
Speaker 2 (38:27):
Tyler Price records.
Speaker 3 (38:29):
So it's just scraping stuff that you've already talked to it.
Speaker 2 (38:32):
It's not even trying.
Speaker 3 (38:33):
Yeah, all right, well you.
Speaker 4 (38:35):
Need to make the prompt more exacting. Rule out anything
Tyler has talked or written about. Yeah, give me something
he doesn't know and try GPT five in the pro mode.
And I think it will succeed well.
Speaker 3 (38:47):
So on this note. This is something I've been asking everyone,
but like, what is an example in your mind of
a really good prompt or one that sticks out to
you that has generated something that you know, maybe you
didn't expect.
Speaker 4 (39:00):
You know, dwark Esh Patel once wrote a very good
prompt and he shared it with me. When I interview
some podcast guests, It's really a long prompt. It's in
hundreds of words, and it asks what questions should I
ask them? Then it goes into great detail. It should
be a unique question, it should be a question they
were not asked anywhere else. Then it's given what you
think their answer might be, and what would be my
(39:22):
follow up question? And yeah, give you give me some
cases where you think their answer might be wrong. And
it goes on and on, and you run that through
the very best models. I think you get good results. Yeah.
Speaker 2 (39:34):
No, I've you know, I've run transcripts of the podcast
before and I say, like, well, where should I have
pushed the guests harder on? What were the weak answers
that they what were the inconsistencies that the guests had
over the time? And I've found it to be a
very useful exercise for things like that. So I do
think like that's the thing which is first it is
(39:56):
it is objectively impressive on many of these from and
I would say objectively useful if you do sort of
a detailed prompting. I'm just curious, what do you what's
your like as a professor from the professor perspective, what
do you think is the right way to think about
how students will be using chadib I mean, I know
(40:17):
that there's a million opinions in academia about well, what's
the right way to test, now, what's the right way
to deal with essays, et cetera. How are you thinking
about some of these challenges.
Speaker 4 (40:28):
We should devote one third of all higher education to
teaching students how to use AI, and right now that's
close to zero, so we don't have the faculty who
can teach it. As part of the problem, often the
students no more than the professor.
Speaker 2 (40:43):
Yea, I'm sure.
Speaker 4 (40:45):
You need to restructure ratically what we do because future
work will be done with AIS, so that's the thing
to teach.
Speaker 2 (40:52):
But like so, just to play devil's advocate, like intuitively,
I still feel like there's value in long periods of
time cut off reading where you're not looking at devices,
where you're training your body to sort of be disciplined
and pay attention and focus. I still think memorization of
(41:13):
facts and numbers and dates and places and names is
very useful in actually having them in your head, et cetera.
Does that seem right to you or is that sort
of retro thinking on my part?
Speaker 3 (41:26):
No?
Speaker 4 (41:26):
Strong agree? And most of all writing, But that's the
other two thirds of higher ed right, I said, one third?
Speaker 2 (41:31):
Oh yeah, so tell us about the other two thirds.
Speaker 4 (41:34):
We should with or without AI just teach students much
more and much better how to write. Most people can't write.
Writing is thinking we should do much more to teach
writing and test writing, and now with AI, that has
to be face to face in a controlled environment where
people are just going to cheat, So that we should
have doubled down on to begin with.
Speaker 2 (41:55):
Yeah.
Speaker 4 (41:55):
So that and just numeracy and basic issues like how
to manage a portfolio, what kind of mortgage to take out.
There are classes that cover those things, but I think
they ought to be front and center of any curriculum.
Basic finance, basic life decisions like how to choose a doctor,
how to prompt the AI you know, for diagnosis whatever,
(42:16):
are relatively neglected in a lot of education. That to
me just seems crazy.
Speaker 2 (42:21):
I want to.
Speaker 3 (42:22):
Ask one market question before we go, which is obviously
there's a lot of talk about an AI bubble at
the moment, and I think the concern from a lot
of people is when you start talking about in new
technology as revolutionary, when you start talking about how you
know the effects are basically going to be infinite, and
the market size is hypothetically the entire world, there's a
(42:44):
risk that expectations overshoot reality, right and we have seen
some people voicing their worries about that right now, and
a little bit of nervousness creeping into the market. In
terms of valuations, Where do you stand on the AI bubble?
Do you see signs of frauth or do you think
most of the CAPEX spending is justified at this point?
Speaker 4 (43:04):
I don't like the word bubble. I would point out
that tech sector earnings are exceeding tech sector capital expenditure.
This is not mostly debt financed, so we're in less
trouble than many people think. It wouldn't shock me if
a lot of these efforts lost money. That was the
case with the railroads, the case with the internet, case
with most things humans have done. But I think it
(43:27):
will endure. It's not like pets dot com, where the
thing just gets swept away. These are incredibly well capitalized,
highly skilled companies where the CEOs and or founders are
quite committed to doing this and they're going to see
it through and they're going to succeed. But does that
mean every share value will go up or Nvidia ends
up being worth ten trillion dollars. I don't know. I
(43:49):
wouldn't necessarily predict that. There's always ups and downs, but
this is clearly a very useful thing and we're going
to we as Americans, we're going to make it work.
And we're way ahead of the rest of the world.
Like three quarters of all AI compute is in this country.
That's incredible. Worry what percent of the world's population six
or I don't know, but way smaller than three quarters.
Speaker 2 (44:11):
GPT five on Thinking Mode says you should listen to
Michael Gulisian, who does dream like acoustic guitar using an
open tuning, and it said that given your affinity for
guitarists like John Fahey and Leo Kottke, you'll appreciate him.
Speaker 4 (44:27):
Send me that answer. It sounds excellent. I haven't heard
of that person. Have you very much liked guitar with
open tuning, shure a.
Speaker 2 (44:34):
Monkar, a Hindustani vocal singer. It thinks you'll like sun
oh md Budrul Hawk. I'll send you this list and
Katerina Barbieri modern modular, minimist electronic composition.
Speaker 4 (44:48):
So I will buy some of these already.
Speaker 3 (44:51):
Wait, just how about a human recommendation? Oh, give us
a recommend a human recommendation. Since you said that you
know country is pretty good right now, but I guess
you personally aren't that into it. But have you tried
Orville Peck? I think there's a new album out out
this week.
Speaker 4 (45:06):
I think, what kind of music is it?
Speaker 3 (45:10):
Country? But like a very modern type of country. I've
tried to get Joe into it, but uh, I'm still
working on it.
Speaker 4 (45:19):
I like some country, and I love old country. So
Hank Williams, Johnny Cash, Jed Atkins period.
Speaker 3 (45:25):
It has a vintage toned to it, but with a
modern twist. Try Orville Peck and then get back to
us about which was better in terms of the recommendation.
Speaker 2 (45:36):
Tyler Colln Tyler Collen, thank you so much for coming
on odd Law's long overdue conversation. I really appreciate you
taking your time.
Speaker 4 (45:44):
Great to chat with you both.
Speaker 2 (45:58):
Tracy a lot to pull out from conversation. I think
it's very interesting that early observation he made about sort
of legacy institutions and whether perhaps some of the sort
of lack of revolutionary impact yet is just about that
metabolization process into the types of companies that could theoretically
(46:19):
absorb them.
Speaker 4 (46:20):
Yeah.
Speaker 3 (46:20):
I mean I think that's exactly it, right, So companies
are using this mostly as an add on to existing workflow.
You're not going to get the huge productivity boom until
companies are sort of centered from yeah, which you know
is probably going to take people who grew up with
the technology rather than old people like you and I.
Speaker 2 (46:40):
You and I, you just adopted it.
Speaker 3 (46:42):
The other thing I was thinking about, first of all,
insurers are sort of a pet interest of mine at
the moment, but I do think like they are emerging
as some of the really big winners from a lot
of like I guess, the data saturation of the world
right now and the increased sophistication of analytical models and
things like that, and that'll be very interesting to see
(47:02):
how it shakes out. And you could if you if
you took it very, very far as a sort of
thought experiment, you could start to say that, like, well,
the insurers are going to be a more important actor
in terms of setting social standards and regulations in the future,
because they're the ones with the data, doing all the
modeling and saying like, you have exactly an ex chance
(47:23):
of being in a car accident, and therefore you must
do the following things, right.
Speaker 2 (47:28):
I thought also, like I hadn't really thought about, you know,
subpoena ability. I think it is a very big issue,
but it is interesting, right, Like, it is a little
weird that you and I could have a phone conversation
now if you're under oath and they say, what did
you talk about? Joe? I expected you'd probably tell the
truth unless it's very bad for me, and I would
(47:50):
hope that you would lie.
Speaker 3 (47:52):
That's right.
Speaker 2 (47:52):
Yeah, I hope that you would. I would hope that
you that's your hope. I would hope that you myself
to save Joe. Yeah, I would hope that you would
perjure yourself, but like you know, theoretically could get away
with it. So yeah, well we have an email, there's
no chance. And it's sort of interesting. It's sort of
it's obviously it seems a little bit arbitraryed me, but
it is interesting to think about, like, Okay, can we
have a conversation with these entities? And like why do
(48:15):
we have to leave a digital record? And where are
these going to be stored? And I do think in
areas like health and law, which are obviously not obviously
but sort of intuitively low hanging fruit for productivity gains,
how much have we not seen, just in part because
we're still sort of negotiating the transition process as the society.
(48:36):
What are going to be the new rules and norms
about this stuff, where it's gonna be hows, et cetera.
I think that's actually a sort of very interesting question.
Or space space, This is the.
Speaker 3 (48:45):
Space to watch, yeah speak, Now that I think about it,
we probably should have discussed some of the regulatory framework
around all of this a little bit more, But next time,
next time. But the other thing I've been thinking about
lately is economic statistics. Yeah, in a world that's increasingly
driven by AI, and I know that we had the
big productivity discussion and technology in relation to technology in
(49:08):
like the sort of early to mid two thousands. I'm
not sure that ever actually got settled, but I very
much expect that the AI economic statistics conversation is going
to be even like wackier because I'm not sure how
you do things like quality adjustments for something that like
suddenly comes with its own brain and stuff like that.
Speaker 2 (49:29):
So no, it's gonna be super weird. I did, I
don't know. In my mind, I have like a very
these images of like Tyler getting a tour through the
you know, chatch EPTE offices and the people ask you
was like, well, and him having to explain that, you know,
we're not going to have twenty percent GDP growth at
maybe two and a half or sorry productivity growth, and
(49:50):
actually GDP isn't really that bad of a measured more
or less captures the size of the economy, even if
a lot of Internet things are free. Like some of
these classic convers stations, I would like to be a
fly on the wall for some of those.
Speaker 3 (50:05):
Tyler Cowen in defense of GDP dependent Yeah, all right,
shall we leave it there.
Speaker 2 (50:10):
Let's leave it there.
Speaker 3 (50:11):
This has been another episode of the Odd Lots podcast.
I'm Tracy Alloway. You can follow me at Tracy Alloway.
Speaker 2 (50:16):
And I'm Joe Wisenthal. You can follow me at The Stalwart.
Follow our guest Tyler Cowen, He's at Tyler Collen and
of course check out his podcast Conversations with Tyler an addition,
of course, Marginal Revolution. Follow our producers Carmen Rodriguez at
Carman armand dash O Bennett at Dashbot and Cal Brooks
at Cale Brooks. From our Odd Lots content, go to
Bloomberg dot com slash odd Lots with the daily newsletter
(50:39):
and all of our episodes, and you can chat about
all of these topics twenty four seven in our discord
discord gg slash odlines.
Speaker 3 (50:47):
And if you enjoy odd Lots, if you like it
when we talk to Tyler Cowen and give him music recommendations,
then please leave us a positive review on your favorite
podcast platform. And remember, if you are a Bloomberg subscriber,
you can listen to all of our episodes absolutely ad free.
All you need to do is find the Bloomberg channel
on Apple Podcasts and follow the instructions there. Thanks for
listening
Hosts And Creators
Joe Weisenthal
Tracy Alloway
Popular Podcasts
Las Culturistas with Matt Rogers and Bowen Yang
Ding dong! Join your culture consultants, Matt Rogers and Bowen Yang, on an unforgettable journey into the beating heart of CULTURE. Alongside sizzling special guests, they GET INTO the hottest pop-culture moments of the day and the formative cultural experiences that turned them into Culturistas. Produced by the Big Money Players Network and iHeartRadio.
Crime Junkie
Does hearing about a true crime case always leave you scouring the internet for the truth behind the story? Dive into your next mystery with Crime Junkie. Every Monday, join your host Ashley Flowers as she unravels all the details of infamous and underreported true crime cases with her best friend Brit Prawat. From cold cases to missing persons and heroes in our community who seek justice, Crime Junkie is your destination for theories and stories you won’t hear anywhere else. Whether you're a seasoned true crime enthusiast or new to the genre, you'll find yourself on the edge of your seat awaiting a new episode every Monday. If you can never get enough true crime... Congratulations, you’ve found your people. Follow to join a community of Crime Junkies! Crime Junkie is presented by audiochuck Media Company.
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.