00:00:00
Well, suffice it to say, one of the biggest stories of the last year has been artificial
00:00:04
intelligence, how companies are implementing it within their structure of their firm,
00:00:09
but also how the public is using it as well. Plus, there are so many questions about what
00:00:14
lies ahead. Pleasure to be joined once again by Stefano Pantoni, Wharton professor, to discuss
00:00:20
what we've seen over the last year. Stefano, great to talk to you again. How are you, sir?
00:00:26
Great to be here. Thanks, Dan. Well, let's start right there and just get some general thoughts on how you have viewed what
00:00:33
we have talked about, but what we have seen play out around artificial intelligence in the last
00:00:38
year. Yeah, so this is the end of the academic year, a good moment to reflect back on the last
00:00:45
few months. We've been meeting every month since last summer, and I was giving some thoughts to
00:00:51
what were the biggest stories over that course of that year. There are many, actually. One could
00:00:57
spend a lot of time dissecting that, but let me give you maybe three top line ones. The first one
00:01:03
is that the models keep getting better, and there were a lot of discussions around last summer on
00:01:08
maybe the progress slowing down and maybe those scaling lows on training, meaning bigger models
00:01:16
performing better were running a bit out of steam. In fact, what we've seen is that the models have
00:01:23
continued getting better. A lot of it is because of those scaling lows in inference that we've
00:01:29
discovered, where the same model is going to perform better when you make it think harder,
00:01:35
and so we are still seeing rapid improvements on all kinds of metrics. So that has been,
00:01:41
I think, the biggest story. Many people were not expecting that, necessarily. The second one has been, over the last six months, really the taking off of agents. There have been
00:01:53
discussions about agents already for over a year, but for many months seemed to be more of a promise
00:01:59
or a speculation. But they're really here, and they are here now in terms of coding agents.
00:02:06
They've been taking over basically coding almost everywhere, and you've seen the rise of personal
00:02:12
agents with global bots and lots of other initiatives by many companies. So that's really
00:02:20
another big story. And maybe the third pillar here is that there really continues to be two races in
00:02:27
AI. There's a race at the high end, the top US labs investing massive amounts of money to build
00:02:35
bigger, more sophisticated models that keep breaking all the benchmarks. But there is another
00:02:40
race, which is to make models more efficient and for smaller models to perform better and better.
00:02:49
And so those are races around model efficiency, about maybe open weights models, and here's where
00:02:56
China has seen also very strong improvements. So those will be my three top line stories for the
00:03:03
year. So let me follow that up by asking you kind of the role that companies are taking right now
00:03:11
in terms of the use of AI and how they expect artificial intelligence not only to improve
00:03:18
their operations on a daily basis, but what their expectations are. I mean, I would assume
00:03:24
a lot of people have talked about this kind of being a beginning of the internet moment once again
00:03:29
with AI. Are we, you know, seeing businesses plan five years out already for the use of AI,
00:03:37
or because of how things are continually changing still, is it still not developed
00:03:45
to that point with companies? Both companies and consumers are adopting AI very rapidly.
00:03:52
Companies' investments in AI are booming, and also their costs are booming. In more recent weeks,
00:03:59
a lot of the headlines have been about the token cost, that basically we are going to have to see
00:04:05
huge budget being spent on AI tokens. And the pricing of the AI labs is changing from those
00:04:14
more subscription-based and limited-use sort of subscriptions to pay-per-use, so token-based
00:04:21
pricing. And basically, the usage is exploding. To me, the funny thing here is, last year,
00:04:29
looking back a few months ago, what has changed? Well, last summer, everybody was worrying about,
00:04:35
are we overbuilding? There's this overinvestment in data centers, and so much money is going to
00:04:42
this. Will we ever need all this compute capacity? And we're worried about the AI bubble that way.
00:04:48
And now, instead, we are in a compute crunch, and we haven't gotten compute. And so, everybody's
00:04:54
worrying about that and saying there isn't enough data center that we need to build a lot more. So,
00:04:59
I think that is really a sign of how things have been changing, I think. Let me, again,
00:05:04
we already had comments this week from Masayoshi Son, the CEO of SoftBank, talking about the AI revolution being 50 times bigger than the dot-com boom. When you hear a
00:05:16
statement like that, what do you think? I don't know about 50 times bigger than
00:05:21
the internet. That's, it's hard to know. By the way, SoftBank is now the most valuable
00:05:26
Japanese company, taking over Toyota and market cap. So, certainly, Masa has got something to copy. And in terms of the relative importance compared to the buildup
00:05:40
of the internet, I think it's comparable in many ways. It's difficult to know which one is bigger.
00:05:45
I do think that the arrival of such smart artificial intelligence models is really a
00:05:51
momentous change. I mean, for society, for humankind overall, I think we don't know,
00:05:56
at this sense, we don't know how much better these models can be. But in a way, they don't have to
00:06:00
get actually much better than they already are to change everything, change everything.
00:06:05
And because, let me slide back to the public consumer part of this, as we have seen all of
00:06:13
the different models update, seemingly week by week or sometimes day by day, it feels like that the development of AI is kind of bringing the public along with it, correct?
00:06:26
Well, yes and no. It is bringing the public along. If you look at AI takeoff, AI usage,
00:06:34
I mean, today, over a billion people are using Chachapiti, for example. On the other hand,
00:06:41
one of the big stories of the last few months, I think, has been a growing AI backlash.
00:06:47
We've seen that in many ways, from personal physical attacks on AI leaders or companies to
00:06:54
very strong and negative vibes among young people. In recent weeks, for example,
00:07:01
there have been several commencement addresses in college campuses that have been booed when people
00:07:07
were talking about AI, things of that nature. So I think it's a mixed picture. Clearly,
00:07:12
people are concerned too. So then doesn't that put even more focus on how the regulatory side plays out as we move forward
00:07:19
with all of this, correct? Well, regulation is obviously very important, like it is important in everything.
00:07:25
I don't think that we should expect very strong moves by the federal administration in the US
00:07:32
anytime soon, although I don't know. But I think we have seen a lot of activity and action at the
00:07:38
state level. And of course, the AI Act has become law in the European Union. I think this is a
00:07:49
safety and some guardrails are probably needed. On the other hand, we really don't want to stifle
00:07:55
innovation. And we really still don't understand exactly how this technology is going to land
00:08:00
and what the real concerns are. So we are in a moment of rapid change. Regulation should be
00:08:06
deliberative, should be consultative. It takes time. And so I'm not an expert on regulation,
00:08:11
but certainly this will be an area where we're going to see some discussions in the coming months.
00:08:15
Okay. So then if that's at least one area of concern, are there other things that you are
00:08:22
at least watchful of that you want to be, I don't necessarily want to say wary of,
00:08:27
but at least you're thinking about in terms of potential concerns as AI continues to develop?
00:08:33
Well, many, many things in terms of company usage. I think we have to understand the tradeoffs
00:08:39
between cost and benefit. For example, the token costs, they have not been priced
00:08:45
accurately before because these companies have been running through massive burn rates.
00:08:50
And the moment that people are going to actually pay for the tokens, we're going to have to see
00:08:54
some calibration on what kind of tasks and jobs do we really want to use this technology for.
00:09:02
You've got, I think, exciting development in physical AI. I think it's going to be
00:09:06
one of the most interesting areas in the coming months. So we've seen a lot of
00:09:10
interesting development of the boundary between robotics and AI. And for consumers, to me,
00:09:17
personally, one of the most interesting areas that I find is how AI can basically become almost like
00:09:26
a relationship partner for consumers in different shapes and guises. But how do we deal with the
00:09:32
technologies that have such human-like features? And we've never had anything like that before.
00:09:37
To learn about. Are there areas beyond that, that maybe even further out, that you are very interested to see
00:09:47
how they play out in terms of the impact of AI in the culture? I think there's many interesting things. I can tell you about one project that we're
00:09:57
working on right now. It's still not even submitted for a journal review, but I'm very excited about this. It's led by a postdoc in the marketing department of Wharton
00:10:07
called Ben Lyra. And what we are doing here is to study what we call synthetic contact.
00:10:13
The idea that you might be able to improve intergroup relations by having AI impersonating
00:10:20
a member of the outgroup. And we do it in the context of political polarization.
00:10:25
And we show that basically even brief interactions with a member of the outgroup impersonated by a
00:10:30
for example, a Democrat talking to a Republican bot, is actually creating less animosity and
00:10:37
more warmth towards the outgroup. And people realize that actually these people are not
00:10:43
strange creatures or monsters. Actually, AI can help people empathize with outgroups.
00:10:50
This is just one example of the many, many, many applications that we can think of.
00:10:55
How has this past year that we've been talking and all of the things that have occurred,
00:11:01
has it impacted you as a researcher and a professor in terms of just everything that
00:11:07
has gone on in the last year? There's so much changing. And I think I probably
00:11:13
not saying something that others are not experienced, which is it can feel quite
00:11:19
overwhelming. The pace of change is so rapid. And especially for those of us who are not
00:11:26
the super tech adopter early kind of person tweaking with AI agents on your little Mac,
00:11:35
PC or whatever. If we are not the kind of people, then I think this change can be unsettling,
00:11:41
can be difficult. It's also, it's a strange paradox because on the one hand,
00:11:45
it's incredibly empowering. Those of us who used coding agents, like say, you know,
00:11:50
Claude Cowork or something like that, you find that your capability frontier has expanded so
00:11:57
much. There's so much more that you can do so easily. It feels very empowering. At the same time,
00:12:03
it's also feeling threatening and you wonder about the value of your skills and your ability to keep
00:12:10
up. And so it's this very mixed feelings. But I think overall, I find it an incredibly exciting
00:12:17
technology. And every day I get, you know, experience awe and amazement when I see what it
00:12:22
can do. Stefano, it's been fun talking to you this last year. Enjoy your summer and we'll see
00:12:27
you next school year. Thanks very much. Thanks, Dan. It's been great talking to you all.
00:12:31
Thank you. Stefano Pantoni, Wharton Marketing Professor, joining us here on the show.