00:00:00
Well, every first Wednesday of the
00:00:02
month, we are talking about the world of
00:00:04
artificial intelligence. We obviously
00:00:06
know how important this topic is and
00:00:08
these discussions are at the moment.
00:00:11
Once again, pleasure to be joined by
00:00:12
Stephano Pantoni, marketing professor
00:00:14
here at the Wharton School and
00:00:15
co-director of Wharton Human AI
00:00:18
research. Stephano, great to have you
00:00:20
back with us. Thanks for your time.
00:00:22
>> Thank you. Then,
00:00:23
>> you know, it's funny you noted to me
00:00:24
earlier this week, I know this is
00:00:26
something you've looked at. Uh we're
00:00:28
starting to see some interesting
00:00:29
partnerships kind of play out in the
00:00:31
world of AI. Uh OpenAI is partnering
00:00:34
with Nvidia. We're seeing other types of
00:00:36
deals right now. So you've called it a a
00:00:39
vertical integration going on. What do
00:00:42
you make of this going on and how it
00:00:44
could potentially impact the future
00:00:46
success of AI?
00:00:48
>> Yeah, we've seen just earlier this week
00:00:50
Nvidia announcing a deal with OpenAI
00:00:53
where Nvidia will invest 100 billion. So
00:00:55
we're talking about massive amounts of
00:00:56
capital into open AI and that has been
00:00:59
uh one of many such deals between
00:01:03
companies that are un matched in this AI
00:01:06
space. One that you probably read in the
00:01:08
news recently was Oracle. Um we've seen
00:01:13
tie-ups between the big um cloud
00:01:16
computing companies and AI labs for
00:01:18
example Microsoft and OpenAI um AWS and
00:01:22
Antropic. And a lot of these deals uh I
00:01:24
think you can look at them from a point
00:01:26
of view of vertical integration. Somehow
00:01:28
I could imagine that for example Nvidia
00:01:30
collaborated with OpenAI could help
00:01:32
Nvidia understand how to best design
00:01:34
chips for AI training. So perhaps there
00:01:36
is interesting transfer learning this
00:01:38
way. But some people are also wondering
00:01:40
about a a certain circularity in this
00:01:43
investments where Nvidia will put money
00:01:46
into OpenAI. OpenAI then gives the money
00:01:48
back to Nvidia by buying the chips and
00:01:50
the similar you know um AWS gives money
00:01:53
to Entropic and then Entropic gives
00:01:54
money back to AWS by buying cloud
00:01:57
computing. So uh some people are also
00:02:00
worried a little bit about this um
00:02:02
inbreeding let's say of investments. So
00:02:04
basically the money comes and goes and
00:02:06
kind of changes hands but it stays in
00:02:08
the same loop. Um but uh certainly big
00:02:11
news on that front. Yeah,
00:02:13
>> it doesn't feel like though that we
00:02:15
should be surprised to see all of this
00:02:16
play out when you think about how
00:02:18
seemingly important we believe AI is
00:02:20
going to be to so much of what we do
00:02:23
over the next many years.
00:02:26
>> Yeah, for sure. And if you uh if you
00:02:29
believe the assumption made by many in
00:02:31
the you know in the tech sector that we
00:02:34
are you know not very many months away
00:02:37
from artificial general intelligence or
00:02:39
something like that that will be able to
00:02:41
perform many many tasks then spending
00:02:44
those enormous amounts of money
00:02:45
certainly makes sense.
00:02:48
>> Uh one of the other areas I know you're
00:02:49
focused on is copyright issues and we're
00:02:52
seeing some of these cases play out in
00:02:54
the court. Anthropic being the latest.
00:02:56
They just settled a a lawsuit.
00:02:58
>> Yeah, that was an interesting piece of
00:03:00
news came also a few days ago. We've
00:03:02
seen judge agreeing to a uh settlement
00:03:07
that um um Anthropic basically um is
00:03:12
making with a set of content creators
00:03:15
for uh um the claim that Antropic has
00:03:18
used pir pirated books to train its AI
00:03:21
models. They settled for 1.5 billion and
00:03:25
this is the first settlement of this
00:03:27
kind and I think we're going to see more
00:03:29
of this agreements between the I
00:03:31
companies that need data for training
00:03:33
these models and content creators,
00:03:35
publishers, authors who have been
00:03:38
providing uh that data. I think that's
00:03:41
the interesting thing because, you know,
00:03:43
I think the assumption was for for many
00:03:45
many years that the distributors had the
00:03:48
the edge, had the control in this
00:03:50
process, but we're starting to see the
00:03:52
content creators, and we're seeing this
00:03:53
play out in the music industry a little
00:03:55
bit, that the content creators feel like
00:03:57
they have more of a say right now.
00:04:00
>> Yeah. And I think increasingly creating
00:04:02
good uh content that can be used by AI
00:04:05
tools to uh um AI companies to create AI
00:04:09
models. I think it's a very strong uh
00:04:11
value creation um you know potential and
00:04:14
if you look at for example few months
00:04:16
ago Reddit's IPO a lot of the um
00:04:19
narrative around the stock was about the
00:04:22
value of Reddit data for training IM
00:04:24
models. So it's really economically
00:04:25
significant. Uh you guys recently just
00:04:28
finished a Gen AI conference. Give us a
00:04:31
little uh background on what some of the
00:04:33
important themes of the conference were.
00:04:37
>> Yeah, we hosted our generative AI for
00:04:39
business conference uh three weeks ago
00:04:42
in San Francisco. the third year running
00:04:44
we're doing this and it's a great way of
00:04:47
bringing top researchers working at the
00:04:50
uh cutting edge of AI applications in
00:04:52
business over to our San Francisco
00:04:55
campus to discuss AI uh for a couple of
00:04:58
days lots of interesting stuff to talk
00:05:00
more but we don't have the time but just
00:05:02
to highlight one topic where we've seen
00:05:04
a few sessions there were three or four
00:05:06
sessions where researchers were
00:05:08
exploring the power of synthetic data
00:05:10
where basically the idea would be can We
00:05:13
learn something useful about people by
00:05:15
talking not to people but to AI models.
00:05:19
And the idea will be predicated on the
00:05:21
fact that these AI models are trained on
00:05:22
so much human data. They know so much
00:05:24
about how people behave and how people
00:05:27
think that you might be able to get some
00:05:29
useful insights about for example your
00:05:31
customers, what they like and how they
00:05:33
think about the product category um by
00:05:35
just interrogating those networks rather
00:05:38
than uh going to talk to real consumers.
00:05:41
maybe you know in addition I would say
00:05:44
but we've seen quite a few interesting
00:05:46
use cases and we are just learning how
00:05:49
this is going to work but it's a very
00:05:50
interesting new frontiers in the uh
00:05:52
industry of marketing services and
00:05:54
customer insights
00:05:56
>> and I'll finish up I know you had maybe
00:05:58
wanted to touch if we had time on a new
00:06:00
study that had come out of Columbia
00:06:02
University as well
00:06:04
>> yeah that was one of the studies being
00:06:05
presented at the conference was really
00:06:07
interesting they uh um a team of
00:06:09
researchers led by Olivia Tuba has
00:06:11
created a uh digital twin data set where
00:06:15
they are interviewed interviewed for
00:06:17
several weeks, 2,000 people and then
00:06:20
they uh created a model of these people
00:06:22
based on all the answers that they got
00:06:24
on over 500 questions and now you can
00:06:27
have a synthetic version of this panel
00:06:29
that you can use to run studies, ask
00:06:31
questions and so forth. So it it would
00:06:34
be used just specifically for that
00:06:35
because it seems like a very unique kind
00:06:37
of development and how it could
00:06:39
potentially be used uh larger scale in
00:06:41
the public moving forward.
00:06:43
>> Yeah. And lots of researchers are
00:06:45
interested in using this kind of data to
00:06:46
test the capabilities of AI exactly in
00:06:49
this domain of synthetic data.
00:06:51
>> Stephano, always great to talk with you.
00:06:53
Thanks very much for your time this
00:06:54
week.
00:06:55
>> Thank you Dan. Stephano Pantoni,
00:06:57
marketing professor here at the Wharton
00:06:58
School and co-director of Wharton Human
00:07:01
AI research.
