CASSIE MCGRATH: Hello, and welcome. This VPAL Signature
Event is brought to you by the Office of the Vice
Provost for Advances in Learning in partnership with the
Harvard Alumni Association. Thank you all for
joining this afternoon. Join our panel of esteemed
experts for an exploration of the future of generative
AI and its implications for education, work, and society. Our panelists will share
their expertise on some of the latest technological advancements
and explore their implications. They
will also examine the
ethical considerations of AI and how it can be used for good. And now without further ado,
Bharat, the virtual floor is yours. BHARAT N. ANAND: Thanks, Cassie. Thanks so much. Good afternoon, everyone. Thank you for joining us. We're going to jump
right into this topic. So I will say rarely is there a topic
which attracts so much attention and about which we know so little in
terms of how it's going to play out. But we're very fortunate to have
colleagues today who join us
who will dare to speculate
about the future, but also bring relevant
expertise to the table. All four of my colleagues
are folks who've worked with AI tools and algorithms
in different though related areas. All are colleagues at Harvard. All have been gracious enough
to spend this afternoon with us on the topic of generative
AI and its implications for a whole variety of things. And we'll touch on four. Research, education, work, and society. So thank you, Rebecca,
Dustin, Latanya, Vijay. Let me
just start by trying to
level set around terminology. And I don't know who
wants to take up this one. Just to set the language straight,
there's generated AI, there's ChatGPT, there's GPT-3, 4, there's open AI. What's the difference
between all these things? Vijay, do you want to take a crack? VIJAY JANAPA REDDI: I'd say that
generative AI is the key word, right? ChatGPT is an instance of generative
AI specializing in natural language processing as an example where you type
and you're able to a
ctually converse with this agent. But talking about even
just ChatGPT as an example, there are multiple versions
of that sort of an agent. ChatGPT-4, which just came if
you haven't played with it, is the ability to understand
images, text, and so forth. And there are multiple
models of information that it can actually handle, whereas the
old version understands just raw text and does NLP. I'll stop there and let
others chime in too. BHARAT N. ANAND: Anyone would
like to add anything to that? DUS
TIN TINGLEY: Yeah, I also think
it's important to understand that there are a lot of companies
getting involved in this space and have been involved for a long time. OpenAI is the one that's been
in the headlines recently, but Google's been a player in this
generative context for a while, and in ways that you have
probably already benefited from, but just didn't know. So sometimes you'll see
if you're using Gmail, it will suggest how to complete
your sentence, which I love, because I hate typing
. So it's been around. There are lots of companies
that are using and building out these types of capacities. OpenAI's just one of them. And obviously, they've been
partnering with Microsoft and whatnot. So it's a big and
exploding space and a lot of companies that are trying to
figure out their business models. BHARAT N. ANAND: Just to
summarize what both of you said, the distinction is between companies
of which OpenAI is one example. The software model, so to
speak, AI-driven, which is GPT-3,
GPT-4, different versions. And then ChatGPT is sort of
this dynamic chat bot, right? It's like the Google
search bar and search as opposed to search more generally. OK, that's great. What I'd like to do is actually, there's
three questions as part of a short poll that Cassie, if you can
just post in the chat, we'd love for all the participants
to fill out, just to get started while we engage in conversation. All right. So for the panelists-- and
by the way, this link in chat will take you to so
mething
called Poll Everywhere. And there's three questions in
that poll if you can fill it out. So back to the panelists. Each of you has been doing work in
AI and related areas for a while. And yet, there seemed to be sort
of a big deal moment last summer. And I'd love to hear from you,
what was that big deal moment, and why was this such a surprise? Dustin, I remember you came to
me in the corridor and said, there's something different here. Can you just share with us? And then I'll turn to L
atanya
and Rebecca and Vijay. DUSTIN TINGLEY: Yeah. I mean, I think my
something different even started a little bit before last
summer when something called GitHub Copilot came on the scene. And it was probably a year before. So GitHub is a service where people
like me and Latanya and others can write code and put it up, and
we can have people collaborate. People can make recommendations. And it's been a really helpful service. So they come along and
release something called Copilot that was a
generative
AI, and it was useful where you could tell it the
code that you wanted to write, and it would write the code for you. And that was like an aha moment,
A, because I'm not the best coder. I always put the comma in the
wrong place and things don't run, and then I lose my hair. But it also made me think, jeez. Like, what did they exactly
train their algorithms on? You have to train these algorithms in
order for them to generate content. And it was a real puzzler
about how they did it. And
they got in-- they started to get into a
little bit of heat about, were they using code
that people like myself had put up on GitHub to make
available to the world in order for them to build a product? And that was like, huh. Where is this all going? But then yeah, last summer we started
to see the release of some of these GPT-3.5, et cetera, models. And the breadth of things that it was
able to do was just sort of stunning. And I actually pulled together
a little mini working group that Vijay
was part of where we were
like, what are the implications of this going to be for education? And then Lo and behold, around
December The New York Times and others started to pick up the advances,
and it's really unfolded fast BHARAT N. ANAND: Thanks, Dustin. Latanya? LATANYA SWEENEY: Yeah, I
would-- well, first of all, I would add to the notion
of generative I think is what really creates
the excitement around it. So AI algorithms were
doing decision making. That was the way before. And people w
ondered what it was
trained on and whether or not that training data had introduced bias. And there were lots of great examples
of that and lots of stories around it and lots of questions around the
use of this sort of automated tool that's making these statistical
inferences and our use of it. But what made generative
different is it creates-- it seems to create something. Something that didn't exist before,
all of a sudden we get a new one. So generative photographs,
generative videos. All of
these have been pretty stunning
because in some ways as humans, we think ourselves are
the only ones creating. And it's not creating
in the way humans are. It's a kind of mosaic regurgitation,
but in a coherent manner. And that's why it
really made the moment-- I think is making this
feel really different. It's around creativity, and
it seems to be creative, but it's really building
off the creations of humans that came before it. BHARAT N. ANAND: Hmm. Thanks. Rebecca? REBECCA NANCY NESSON: Hi.
Yeah, I'd just start by
echoing what Latanya said. I think one of the things
that was really striking about it is in the coherence and the
amount that that gave me and lots of us pause about the ways that we normally
evaluate authenticity or reliability of things that we're interacting with. And that when you-- it's hard to
tell whether something is generated. It suddenly calls into question
all of our systems for that. But the thing I actually
wanted to share was for me, my surprise moment came
when my
daughter was playing around with it, and it was funny. And I was really surprised that
ChatGPT was able to be that funny. She asked ChatGPT to generate an ad for
emotions, an ad jingle for emotions. And it generated, attention, all humans. Are you tired of feeling the
same old emotions day after day? Are you ready to add some excitement and
variety to your emotional repertoire? And then later on says,
we've got everything from tried and true favorites
like happiness and love to the more
complex and nuanced
emotions like envy and guilt. And it goes on in this vein. And I just thought it was kind of making
fun of humans and all of their emotions throughout and was surprised by that
kind of cleverness or contextual understanding that I
had never experienced before with a generative agent. BHARAT N. ANAND: That's so interesting. Vijay, you've been working with
Google also for a while, right? And was there a surprise moment
for you, or how did you react? VIJAY JANAPA REDDI:
So I ta
ught a course, which doesn't help me with,
and Bharat, you know all about, which is like a tiny amount. So that was a very exciting thing that
I was doing during the COVID time. And then the course was quite popular. And then I was like,
oh, I have to write-- I should write a short
little book on this. And then the time is so limited,
I was feeling very strapped, and I kind of could feel like,
OK, I have to write this stuff. But I don't want to write
all the boilerplate stuff like what is AI and
all the basics. So I started looking
around for services. This was pre-announcement of ChatGPT. There are other services like Jasper
AI and so forth where I kid you not, I wrote 90 pages of a
book in less than eight hours. And I was using already--
like, to Dustin's point, there are other services that exist like
this that are basically-- they had-- GPT's sort of the access point. It's when I started
doing that I was like, wow it's actually generating
fairly good content. I gave it to my wife.
I gave one page to my wife
and I was like, can you tell-- what do you think about this? And she kept asking me, why are you
asking me to read this one page? And I was like, oh, can
you just take a look at it? Because I was wondering if
she can tell whether it was human-generated or machine-generated. And she could not tell obviously, right? And of course, at that
point, once it started generating the volume of content,
that's when I was like, Oh my gosh. I think this is something big. It was rea
lly the first
time I started thinking, oh. You know that experience that you feel
when something transformative-- like when the iPhone shows up at
your door or when you figure out that, wow, there is this
notion of the internet or that the world wide web
is accessible to everybody? When I started using that
service, that's when I was like, this is that next big breakthrough
that's going to change it. And you know why? It's not about-- honestly, I don't
think it's about what it generates. It's mu
ch more that
suddenly there is AI that's accessible for everybody in a
way that they can interact with. Instead of just hearing
about all this AI, AI, AI and its enterprises that are doing
stuff, you can actually do stuff. I think that is when I
was like, oh my gosh. This is going to be incredible because
the world is going to use this in ways that we cannot imagine today. BHARAT N. ANAND: That's fascinating. So if you're using it
to write your books, I guess this has big
implications for whethe
r we allow our students to use
it to write their essays, which we're going to
turn to in a few minutes. By the way, my own story was-- actually had to do with a
mistake that ChatGPT made. And this was back in December. And this is going to
sound a little nerdy. I asked it about whether
there are externalities associated with the Pigouvian tax, OK? This is like econ jargon. And it gave me an
answer which was wrong. And I prompted it back and I said,
your answer is wrong because of x. And instanta
neously
it said, you're right. Here's the correct answer. And it gave a much more
detailed description, which was far more elaborate
than what I had given it. And that was just like, wait. What's going on here? I mean, the ability to
actually change on the fly and correct itself
was pretty incredible. OK. I'm just going to show the poll screen. So it seems like Poll Everywhere is
overloaded already with responses, but we have a good distribution here. So roughly that's what we are
seeing, which
is about 80% of folks who've tuned in have
spent somewhere between zero and 10 hours on ChatGPT. And then we have-- and then we have some on the other end. All right. Let me turn to-- I want to go through three
or four major topics here. And obviously, this has implications
for virtually everything we're doing, but I want to focus on a few areas. Let me start with the implications
for the research enterprise. We spent a lot of time
talking about teaching, but I want to start
with research itself
. So Dustin, I'll start with you. So GPT-3 and GPT-4 are
basically state-of-the-art-- what we call natural
language processing models with something like
175 billion parameters. You actually spent the early part of
your career using natural language processing models in your research. Can you just tell us what
it is simply, and why was this such a big deal
in terms of the evolution? DUSTIN TINGLEY: Sure. So let me put this slide up just
to work people through something. So I started working on
n
atural language processing with a goal to be able to
detect patterns in text. So for example, if
people explained why they voted for one candidate
versus someone else, what were the sort of general
themes that might come out of this? And so something like topic
models was one area that I worked, and that's just asking
this question, like, what words simply tend to
co-occur with each other? So for example, if we were to take
each speaker, what they're saying, and ask, all right, what are the
diff
erent topics that they're talking about, I might talk a little
bit more about some of the algorithms, maybe some of the business models. Vijay might talk a little bit about
some of the more technical things of how they work, et cetera. But it was really about how words
just tend to co-occur together. Or something like sentiment
analysis, which was saying, hey, I know which words are negative
and which words are positive, and let's just get a
general sense of the room. Like, are people feeling
go
od or not, right? And I should say for us,
I got a lot of citations for developing software packages
that did this sort of stuff. My colleague Gary King has been
a leader in this area as well, just providing these sorts of tools. So then you started to see
more advanced things come up. So people started to do research on
something called word embeddings. And this is this idea that
maybe we can figure out words that share a semantic
meaning even though they're different from each other. So for ex
ample, how does an
algorithm know that Massachusetts is similar to New York? Well, they're similar
because they're both a state. And so companies and researchers
started to do work training models to figure that out. And then imagine a version
of that where we're able to do the same thing not for
words, but for whole sentences or whole documents. So there's a lot of really
great work that happened there. And again, this was stuff that as
a researcher, a white collar job, I guess, I was able to h
ave some impact. But now we're in a whole
other world, right? We're in this world of
multipurpose, generative AI. Here we're talking about text, as
Latanya and others illustrated. And here, we're doing
things a lot differently. So sentence completion. Now, we had a version of that earlier. We're style editing, all right? Like, Dustin, don't write
in the passive voice, right? But the flexibility is just transformed. So now it's, hi, write code in language
X to do Y. Or better yet, here's some cod
e in some language. Now translate it for me
to some other language. Or here's 50,000 pages of documents
that I've collected in some, say, sort of legal discovery process. I want you, the algorithm, to
summarize coherently or even be able to have conversations. And the sort of-- where
this has taken research is going in all sorts of new
and fascinating directions. I'll give you an example
that just came out yesterday. So a group of researchers--
again, I'm a political scientist, so I'm going to u
se an
example from politics. What they did is they asked ChatGPT to
compare pairs of legislators and say, which one is more liberal and
which one is more conservative? And then they did that with lots
and lots of pairs of legislators in the US Senate. And by doing those sort of pairwise
comparisons a bunch of times, it came up with a scale
of who's the most liberal, who's sort of in the center, and who's
sort of extreme on the right, OK? They then correlated that
scale that it came up with just
by querying this tool with other
scales that people have developed using how people actually vote. And the algorithm did not
have access to that data, and the correlation was stunning, right? And so it's just an example of how
the research process is not only in a sense replacing, if you will,
or augmenting earlier work that had happened, but now
just opening up new ways to be able to extract information
and make meaningful measurements. So the research phase, we're just
tapping the surface of t
his right now. It's going to move fast. And that's literally just
in a very small domain of me talking about politics. But you know, Latanya has
domains that she's exploring. Vijay-- I mean, on the technical side
it's moving fast, so really exciting. BHARAT N. ANAND: Well, thanks for that. That's really interesting. Just to summarize, so there's a
few things in what you're saying. It's in effect a super smart
sentence completion software. It's not just pre-programmed, right? It's dynamic. It's o
n the fly. And now as Vijay was saying, it's open
to the public, and at least for now for free, right? That's essentially
what we're describing. There's a question in the--
and by the way, folks, please chime in questions
in the Q&A. There's a question about, what
are the limits to this? Which I want to get to in a second. But Dustin, you were
sort of, I guess, half joking that in the early part of your
career, you got a lot of citations partly because you had built up
comparative advantage with
using these kinds of tools. Is that comparative advantage
now basically commoditized? Am I getting that right? Or what are you implying? DUSTIN TINGLEY: It's
getting awfully close. I mean, I don't feel like I'm going
to be able to compete in that area, and that's OK. Put me out to pasture, I guess. But now it's going to be incumbent
on me to be creative in coming up with new ways to apply these tools that
these large companies are coming up with. That is going to be where I
will have a research
advantage. And the example I just gave
from some colleagues at NYU is just one such example. But it's going to put pressure on
me to bring their creativity of how to apply these tools, whereas
before I was the person that was kind of making the tools available. Like, here's a knife. Here's a spoon. Now they've given us a giant tool kit
loaded with lots of utility knives that can do all sorts
of different things. BHARAT N. ANAND: Just to
complete that thought, Vijay, where is most of this innova
tion
now or the advances coming from? I mean, in some-- by the way, I have to
say what Dustin said is quite sobering, because we often talk about
AI and blue collar jobs. This is very sophisticated
skills like the ones he and others have that essentially
now look quite different. But one of the things
you were talking about is most of this is not coming
from the research enterprise within universities. It's coming from companies. Can you just share more on that with us? VIJAY JANAPA REDDI: Yeah,
it's a little disturbing. If you look at some of the fundamental
building blocks of what makes, for example, things like
ChatGPT really work, those fundamental building
blocks are really things that-- like the "Attention is all you need"
paper or like the BERT masking methods, which are all kind of the critical
pieces that kind give rise to the underlying architecture that
does it, you look at those innovations, those innovations-- I hate to say this-- really
came from industry labs, right? And
I think it's getting really hard
for us academics being really good and being really
passionate about this topic to have real technical impact in here. And in part, that reason is
because you need massive-- I'm talking about massive
computing infrastructure to be able to do the
kinds of experimentation that you need to do to
be able to figure out, what are those nuances and knobs
you want to turn here and there? I think that is something that
is going to really hold us back. Can hold us back. I
think it's a concern that
definitely keeps me up at night. I'm like, how am I supposed
to do this innovation? You already kind of
mentioned, Bharat, like, oh, I tend to work with the
companies and so forth. I'll tell you why. Because if you don't, you
will be left in the dust. I guarantee you. It's a really critical thing. And there's so many
of my colleagues just in the computing field
will ask, oh, how are you able to know this and that and that? I'm like, well, it's
because for better or wor
se, I do have to work very
closely with them because they do have access and knowledge to
things that we are still not aware. And by the time they come out,
especially now I can tell you, given the competitiveness
that's in the field right now, doors are closing really fast. They don't want to necessarily
publish all this stuff. Now suddenly, there's
this thing like, OK. My gosh. This is going to be
make or break for us. So that's the big concern
that I have is that academia is struggling in thi
s sort of space. BHARAT N. ANAND: Fascinating. I mean, it's sort of somewhat-- Gary King, who Dustin was
referring to, has also been talking about
the major trend, let's say, with data analysis for the last
20, 30 years where 30 years ago, we were basically buying data sets, right? Or using government data sets. I mean, that was the norm. And now most of the data that we want
to analyze sits within companies. And what are the implications
of that for research? There's several questions
already s
tarting to get at issues around
plagiarism, cheating. Rebecca, I want to start
drawing you in into what the implications are for education,
which are potentially profound. In fact, somewhere near-- and this is my rough estimate-- 50% of the articles
written on ChatGPT are talking about the
implications for what's happening with teaching and learning. There was an Atlantic story titled
"Will ChatGPT end high school English?" which talked about this as basically
the combination of the printing pre
ss, steam drill, and
the light bulb having a baby in terms of the innovation. You were part of the team
overseeing undergraduate education at Harvard College for several years. You actually taught a
course in Second Life, if I understand, more than a decade ago. And now you're Dean of Academic
Programs at the Engineering School, so you've seen the spectrum
from writing to coding. What are the areas of teaching and
learning most likely to be impacted? And if I can start with a concrete
question r
elated to ones in Q&A, which is, what's your view
on cheating versus teaching? REBECCA NANCY NESSON: OK,
that's a lot of questions. I think maybe I'd like to start by
saying that what Dustin said about what he needs in order to be competitive in
his field is to be able to creatively use these tools is, I think, the
fundamental challenge for all of us. If we don't transform
education in a way that it helps our students be ready to do that,
then we're not succeeding as educators. And so at base, I
would say we're
making a mistake if our attempt is to focus on how to ensure that students
are still doing the things that they used to do and not using these
tools when the tools can be helpful. You know, but here's a cynic's view
about the cheating side of things. I teach a large undergraduate class. I have 180 students. I spend a lot of hours
every week on this class. And in the cynic's view
of things, if the students can do the work with
ChatGPT, I might look at it as that I am essentially
teaching GPT
180 times to do the work of this class. And then I would do that
same thing again next year, and that is really
accomplishing nothing. I think the minimal view of it is that
we already operate with an honor system for the most part for our
students where we say, you cannot use outside resources,
whether they be ChatGPT or a person who took the class last year and saved all
of the problem sets that are somewhat similar to the new ones. And we feel OK about
that because we know that t
he students for the most part
have a strong intrinsic motivation to learn what's being
taught in the class if you've done a good job setting
up what they're there for. And that's OK. I think that's a totally reasonable
policy for a faculty member to take. But what's really exciting
is to think about, what can you be doing with the
new landscape that we're in? What should you be doing
with the new landscape? So if you think about
the question of, what is the difference
between a student armed wit
h ChatGPT who has taken
your class and a student armed with ChatGPT who hasn't taken your
class, at the end of the semester there should be something that's
different about the one who did take the class. And where I come down on
it is that first of all, you're much better
off if you're teaching your class in some kind of active
learning, highly interactive format. That is something that the students
are going to be highly engaged in doing and that they, just as a matter
of course in the activit
y, can't substitute a software for. And it will engage them
in the process of thinking about the ways in which
they can effectively collaborate with the software, not just
substitute the software for themselves. And secondly, it will
help us to learn where the real learning is in our classes. So I certainly would encourage faculty
to think about how to incorporate it rather than how to stamp it out. BHARAT N. ANAND: Yeah. I mean, just to pick up
on the last set of points you made, Rebecca, it's
really,
really interesting that in some sense, this process of
adaptation to technology-- and how can we lean in? I mean, we saw this
partly during the pandemic where the notion of take-home,
open book assessments became essentially a
necessity, [LAUGHS] right? I mean, one question is, what
happens to that with ChatGPT? But the broader point is,
how do we think about, A, the nature of assessments,
and B, as you're hinting at, how do we think about what we
teach and how we teach, right? I mean, t
hose are pretty fundamental--
those are pretty fundamental questions because they start getting at what
we're doing in our classrooms. Latanya, you've been using this,
I understand, in your classrooms with your students? LATANYA SWEENEY: Oh, definitely. From everything-- one student
took my final paper assignment where they had to write a
research paper and asked ChatGPT to write a
research paper for them, and literally just gave
it the entire assignment. And it literally wrote a-- it produced a
final report
that was really pretty good. And then another student asked ChatGPT
to write a research paper written by Latanya Sweeney on data privacy. And it literally wrote a research
paper that sounded great by me on data privacy, only I never--
it's an experiment I never did. As far as I can tell,
no one else ever did. It had citations that don't exist. It was pretty amazing. So it's amazing on two levels. One, that it could
produce something that was such a great imitation of something. And
it's that notion of that imitation
that allows us to use it as a tool, as a starting place. But it makes up stuff all
over the place, right? It just adds in, I need
something that looks like this, and just puts it there. And so that's-- so it creates
a lot of false information. And so even if someone were trying
to provide the output of ChatGPT in an academic setting,
the problem with it is you've got to go through and clean
up all the things that are inaccurate, right? And sometimes it's a bit
repetitive. On the other side, the
imitations are so good that when we did lots of experiments where given
a paper, a research paper to a person, one actually that came from science and
one that didn't, that ChatGPT wrote up and we asked it to write
it in the specs of science, and people couldn't really
tell the difference. That's pretty amazing. Even educated people in their field
couldn't necessarily at first glance tell the difference. And then in other writings,
people can tell the differen
ce. But where it really seems
to be almost impossible is when it puts on a
personality in chat, because then as humans we personify it. We believe it's often will-- people will
believe that the ChatGPT, for example, is the person, and the person
who is speaking is not. So I found that was a
really stunning example. BHARAT N. ANAND: Fascinating. I'll share one thing
that I tried, which is I gave it a prompt asking, what should
Harvard's strategy in online education be in a way that preserves the
magic of the campus experience? And you know, it typed out a fine
essay, but it was pretty generic. I then cut and paste the essay
and gave it back to ChatGPT and I said, if this was the answer
to an essay prompt like this, how would you grade it? And it said, I would it a grade
of C plus because it's not Harvard-specific enough. LATANYA SWEENEY: [LAUGHS] BHARAT N. ANAND: Which
was absolutely fascinating because I was like, actually, this
is not just for giving answers. It's for grading answers.
And there's a certain
degree of self-awareness, or dare I say, humility here. LATANYA SWEENEY: [LAUGHS] BHARAT N. ANAND: Rebecca, let
me just cycle back to you. So you have instructors
who are basically saying sort of the equivalent of
laptops down in the classrooms. You can't use ChatGPT. You have others like Latanya
basically saying, embrace it. Use it. There are imperfections. You will figure it out. As an academic leader,
how do you set policy? REBECCA NANCY NESSON:
Well, we do leave it lar
gely up to the faculty in
their classes to determine what their policies will be. We ask all the faculty to be really
clear about it on their syllabus exactly what they expect. But we do find that it's
confusing for students that the policies are so different
across different classes. And I think that's-- I think that's the better
option in the trade-off than to try to have one fixed way
that we expect all students to do it. I also think the challenges are really
different for different classes.
So just back to what Dustin said at
the beginning about GitHub Copilot, early on with that tool David
Malan was using it in CS50. And rapidly, as soon as it came out,
tried it out with the assignments. And not only was it
generating correct code, it was generating the
solutions to the problem set, because I mean, there was a question in
the Q&A about, when is it plagiarism? I mean, in that case I
think it was plagiarism. The AI was actually copying something--
solutions that it had found online
and spitting them back out. And so you need in that
case to be able to have a different policy in
that class and what you're working on in that class
with a particular tool than, say, in a different one where, for instance,
you're working on essay writing. And even a content-based
class with essay writing might have a different policy
than a class that's really trying to train students to write. So I just don't think we're in
a one-size-fits-all situation. What I would strongly
suggest to facul
ty is that they incorporate some forms of
assessments that are a little bit like laptops down
assessments, things that actually happen in class
that are more interactive and can't just be done by a software. BHARAT N. ANAND: I mean, you're
highlighting a fundamental trade-off here, which is in some sense
you want local innovation and local efforts by
faculty to try and see what works and doesn't
work in their classes. The trade-off, of course, is for
students, that can be confusing. And how do w
e balance that,
at least in the short term? Let me just say, there's a whole
bunch of questions here around, how might we address things like false
information, misinformation, the notion that most of the train data for
ChatGPT has been in English? How do we incorporate other languages? Latanya, I'm just
flagging those questions because we'll come back to you
with that in a few minutes. I want to pose this
question to all of you before we leave the teaching
side of things, because I mean, this i
s really an important
set of issues for all of us. I'm thinking of the
London cab drivers, right? The black cab drivers, the taxi
cab drivers in London, who-- and you all might be familiar
with the story, right? I mean, they go through
three years of training to basically acquire knowledge
on over 25,000 streets in London, and it's considered the hardest
test in the world by some people. And then GPS comes along. [LAUGHS] OK? And it's like, wait. What just happened here, right? And over the last
decade, it's sort
of been an interesting struggle just to see how they're trying to adapt. If we can fast forward
10 years and if you were to speculate on, what is the
most different aspect of the content and the teaching in your
courses, what would that be? Whoever wants to-- DUSTIN TINGLEY: I'll
go first, and just very briefly because I know
we've got a lot to cover. Relationship-driven rather
than content-driven. BHARAT N. ANAND: OK. Thanks, Dustin. Great. VIJAY JANAPA REDDI: I'm
just going
to say that I don't know about 10 years, people. We didn't know this thing
existed one year ago. And I was talking to a
couple of researchers at Google who work on these things,
and they were like-- they were saying, we didn't know we could
do this kind of stuff. How are you going to plan
for 10 years from now? We don't even know what's going
to happen in the next six months at the rate of which this is
actually happening right now. BHARAT N. ANAND: Yeah. LATANYA SWEENEY: Yeah-- VIJAY JANAPA RED
DI: That
is an important question. Sorry, Latanya. LATANYA SWEENEY: No, no, no. I was actually thinking
even in a shorter timetable. I've been teaching a version of these
courses around technology society classes for over two
decades now as we've been in the middle of this
technological revolution. And this will be the
first time where we're asking the students to start
with a paper that actually is written by ChatGPT on their
particular research area, and then edit from there. That's pretty ama
zing. I never could have predicted that. But everything about this sort of
revolution that we're a part of has already in just
so many-- so few years changed everything about
our lives, even how we're giving this presentation today. Online, how we work, how
we play, and so forth. And we don't know when
the end is going to come. Many historians call it the
third Industrial Revolution, and it's just not clear
when the end is in sight. It's hard to know-- I mean, if you went 10
years earlier from h
ere, you know, where the new
technology was Facebook, you could not have foreseen
where we would be today. REBECCA NANCY NESSON: I
think I would say for me, it feels like a comparison
to the kind of back and forth in baseball between the
pitchers and the batters. And this is a strong
shift for our students towards needing to be able
to read and to understand really carefully and critically
more so than to be able to generate. And so I think the assignments are going
to push in the way Latanya's
saying. It will start from
something that's generated and they need to be able
to make sense of it. BHARAT N. ANAND: Yeah. I think-- I mean, in some
sense, I think one of the things that you are all pointing to is, what
is it we're really trying to assess now? Rebecca, you and I were chatting
yesterday-- what's that? VIJAY JANAPA REDDI: What
you're trying to assess, Bharat, is the ability
to generate prompts. [LAUGHS] Theoretically, it's
called prompt engineering, right? But in fact, I think som
eone in
the chat actually raised that. But it comes down to that, the ability
to understand how to play with this. And one thing to keep in mind, it's not
about just prompt generation, right? It's also understanding the context. To your point, Bharat, like where you
give something and you go back to it. And so it's this interplay
between the system. It's not just a one-shot, oh, I put
in something and I get something out. And I think that's that new generation
that students will all be looking a
t, is like, what is prompt
engineering as a science? LATANYA SWEENEY: I
have a great example-- BHARAT N. ANAND: Sorry. Go ahead, Latanya. LATANYA SWEENEY: I
was just going to say, I have a great example
of prompt engineering. My son-- he's 14 years old-- he
wants to build these card games that have these amazing images on them. So he uses a generative AI
platform to generate these images. The theme of them are
all chess-oriented. So he became this
amazing prompt engineer to create exactly the ki
nd
of image he really wanted. And the images are absolutely stunning. And so there is something to-- but that's us learning how to get
what we want out of this machine. Right? BHARAT N. ANAND: There's-- yeah, no, that's a great point. By the way, just to make sure,
I mean, the flipside of this is where Khan Academy
is going now, right? And Sal Khan announced two days
ago the partnership with OpenAI where in effect, they're
using it as a tutor. So it's not giving answers, but
if I want the answer
to something, it's basically going to work
with me like a personalized tutor to keep prompting me with the questions
so I can figure out what the answer is, which is equally interesting. There's a question around, what are
the skills ultimately that we'll value in terms of writing and coding? Or what are the markers of
intelligence, I guess, more broadly? And Rebecca, we were
chatting about this yesterday and you used the analogy
of math addition, simple math addition and
neat handwriting [LAUG
HS] are things we sort of
don't care about anymore, or at least they're not valued
the way they used to be. Do you have any sense of what
are the kinds of skills we will-- REBECCA NANCY NESSON: I don't want to
shortchange ChatGPT or generative AI, so I don't know exactly what
things it's going to be able to do, or even is potentially able to do now. But I think that some of the things
that seem to me to be valuable are to be able to decide
what problems to work on. Maybe ChatGPT can do that. I'm
not sure. Maybe these models can. I wonder that about Latanya's
example about generating a whole paper written by her. And how interesting was the experiment
that it came up with for you to do? And I also wonder about the ability
to do more transfer-like activities, connecting of knowledge
across domains in a way that's really creative and interesting. BHARAT N. ANAND: That's
really interesting. By the way, there's one
question about, is there a risk for ChatGPT
replacing interactions with the
on-campus
experience from a professor? And I'll just share my thoughts on this. One of the things that was fascinating
to me over the last three years with remote teaching was the
extent to which a number of faculty started thinking about once they
experienced online tools, it's like, wait. This is actually pretty
good for content transfer. And then it forces you
to think about, why do we actually need to
meet in person, right? What's that
synchronous-asynchronous mix? What's that online versus
in-person mix? And you start appreciating the
importance, as Dustin was saying, of relationships and connection and
community, and those kinds of things. And I think this is going to force us
to do exactly the same thing, in effect. I want to turn to
thinking about the broader applications of this in industry. And Vijay, you've been
working on AI integration in a whole range of
applications for many years, and you lead a lab at Harvard. In some sense beyond just individuals
like us interacting w
ith generative AI tools like ChatGPT,
can you say something about the broader range
of applications where we're likely to see this
in business and work? How do you think about that? VIJAY JANAPA REDDI: I
think the big thing is we're all talking about ChatGPT
because it's the hub of the world, it's the poster child. But in reality, from what I've
gathered from many companies, they'll say that, oh, we're
not going to build a chat bot. That's not what this technology's
really going to unlock. That
is just the thing that
people can engage with, and it showcases the possibilities
of what is really possible. And the thing you have
to realize is these are-- the reason they're
exciting is because they're able to synthesize things that
are well across domains, right? So there's interactions
that you and I might not be able to pick up because we can handle
a limited sort of domain knowledge. But this thing can just suck in
whatever connections it sees, and sometimes those
unexpected connections
are what gives rise to all
sorts of new information that you extract from the raw data. So I think ChatGPT is just the tip of
the iceberg in being able to do that. And I think a lot of research is showing
that it's not the kind of service that people really care about. Video content and online media, talking
about blogs and those kinds of things, apparently it turns out
it's like a mass market. Law, I don't know. Latanya, you'd know
much more about this. But I've been thinking, oh,
this is prett
y interesting. I wonder-- I was going to ask you
about this in person, in fact. I was like, oh, if I wanted
something drafted up, would I really need a lawyer
or someone really experienced? I mean, especially with law I feel
like it is kind of so, in a way, rule-driven, perhaps. And I was kind of wondering,
in such industries maybe it's transformative because
it's like suddenly, I don't need to pay, I don't
know, thousands of dollars or whatever it costs to get a lawyer. Now suddenly, the best,
talented
lawyer is probably accessible for me for $0.99 per request. So that's the way I'm thinking about it. Latanya, it'd be nice to
hear your thoughts on those. LATANYA SWEENEY: Yeah. You know, all of us are very-- certainly spoken to
the positive advantages and excitement around
these generative tools. There are these incredible flip sides. And the question that
Vijay's asking starts to begin to think about what
some of the flipsides look like. This idea of false
information and just making
up stuff, how do you deal with that if you
wanted to use it in a setting like law, right? If I wanted to understand
the law, like, suppose it was primarily being learned
off of the actual legal texts, it's phenomenal just
to help me understand. We have 2,167 privacy laws in the United
States, the last time I counted them. And if I wanted to
know, is a certain act OK, it would be fantastic to be able
to just ask that in a generative form like that, just in a prompt. But I can't trust the
answer t
hat I might get right now, or even for a while. So solving that problem is real. The other thing too, when we think
about the future of the internet, within three years I would
guess 90% of the content, maybe even 95% of the content is no
longer going to be generated by humans. It's going to be generated by bots. And these bots are
basically statistically regurgitating variations of
themes of things that humans said and then regurgitating
variations of things that other bots have said or written
. And what does that mean? What does that mean about
our sense of what is real, what is truth, what is knowledge? And these bots are being taught
primarily on data from the internet, in the case of ChatGPT. And it learned on the internet the
internet has all kinds of biases, all kinds of issues,
all kinds of problems. And ChatGPT shares those
issues and problems, as it will often regurgitate things
that it really shouldn't be saying. And right now they have a filter on
it to try to not let those
things out. So there are a lot of issues. And then I think even in the chat
was the intellectual property question that in the GitHub or in the
images, we as humans, we create them. There's a kind of
economics involved in what we think was the
production of this work, whether it's a particular way of
writing poems or a particular art form, or what have you. And then all of a sudden, ChatGPT
appropriates it and uses it and is generating it en masse. How does that work? So it brings on intellectu
al
property questions that we haven't examined before. BHARAT N. ANAND: Latanya,
you've been, I mean, in some sense working for over a
decade on some of these issues, right? Algorithmic bias, privacy,
re-identification of data. As you think about generative AI
models and tools, I'm just curious, how does the scope of
these issues change? Is it incrementally,
or is it a matter of-- is there a discontinuous change? And I guess the more fundamental
question is based on your experience, what can we
do about this? Take just the example
of algorithmic bias. And it'd be great if you
could just explain simply what that means to the audience. LATANYA SWEENEY: Yeah. Well, the way-- I would go
back just one step further. When I started this-- I'm a computer scientist by training,
and when I was a graduate student you could just sort of see
this revolution coming. It promised to transform our lives. And we just-- many of us just believed
there were no downsides to it. It was all going to be upside
. But right away, one of the first
things that I was early to see was the problems with data privacy,
how easy it was to re-identify people, how hard it was to keep your data safe. And so in many ways, that became the
first major technology-society clash. Then as we go forward, the next
big technology-society clash was when AI algorithms were
used for decision making. They would train-- they still are. They would train on
data and then become-- make decisions for us
or advise our decisions like
in criminal justice or credit
worthiness, and so forth and so on. And these algorithms are used
widely, but most people don't even know they're part of making decisions. They can't explain their decisions. And if they're trained
on biased data, they just produce biased results
because they're just replicating statistically
what they were trained to do, even though they look like
there's going to be impartial. So that's sort of the next wave. The third wave has been around
misinformation and disi
nformation and the use of that information,
the spread of it on social media. We don't really know how to
establish truthful content at scale or what trust means at scale
or moderation at scale. We haven't figured any of those
out around the issue of moderation. So those are three major tidal
waves, and now comes generative AI. The time between them is shorter. Between privacy and algorithmic
fairness was one time, but only a couple of years later before we
got to the issues around democracy, an
d only a few years shorter
than that before we get to these issues around generative AI. So these clashes are coming faster. And the truth is, even though
data privacy goes back to 1997, we still don't know the answer. We still don't have the solution. So these things are
transforming our society. We can't even enforce many
of our laws and rules. It's upending others
of our laws and rules, and we are moving at a
snail's pace in comparison. It is definitely an
issue for our society. BHARAT N. ANA
ND: Is there-- well, [LAUGHS] is there a
silver lining, or what-- is there any cause for optimism on
any of the fronts you described? LATANYA SWEENEY: [LAUGHS]
Yeah, well, you know, I'm still that graduate student. I'm still the graduate
student who wants society to reap the benefits of these
technologies without the harms. And many of us-- over time, we are definitely
growing a workforce of technologists who want to work in the public interest,
who are trying to shore up government, trying to s
hore up sort of consumer,
civil society, and so forth. All of the people who
would normally help society maintain itself need to sort of be-- need to have technologists
who can help them do their job in this sort of new area
at the speed at which these issues are coming. And so there's a lot of movement
towards public interest technology. We think that'll make a big difference. Whether that'll save us, we'll see. BHARAT N. ANAND: Yeah. No, and obviously, that's one dimension. I just want to clos
e with one
question that just came up. Are there steps that you believe
tech CEOs can be taking now to roll out safer AI-enabled products? I guess to any of you. LATANYA SWEENEY: I'll just
give a quick one-liner. Two weeks ago, I did have a great
meeting with venture capital firms in Silicon Valley trying to
change the way this cycle is working to help them as part of when
they're developing new companies, investing in new companies actually
think about technology-society clashes, because a lot
of times they can be
avoided simply in the design stage, and it's really easy. It's much harder once these
products come to market. BHARAT N. ANAND: Hmm. Vijay, anything to add there? VIJAY JANAPA REDDI: Yeah. So around safer AI, one of
the issues is that as people are building these models
right now, everybody's really focused on what new
capability they can really unlock, because they really want to
push the state of the art. It is a very competitive sport. And so that's what
people are really
doing. But then we have this notion. We're working with a community of AI
researchers spread across Google, Meta, and a whole bunch of other
companies and academia to build this notion of safe
AI is better AI for everyone. And the intent is that as we use
these generative methods, what we need are mechanisms to collect those
adversarial examples, the cases where these things are breaking down. For example, like ChatGPT or
any of the generative ones, you want to be able to
say, OK, well, this on
e is considered a harmful
image or chunk of text. And everybody has to do that. You and I have to do that, because
the way you think and the way I think are different. So what I might think is totally fine
is something that my wife will totally disagree with. That's totally valid because there
are two different perspectives going into what we're looking at. But if we can capture those examples
with a little bit of context and then we can build these large
corpuses of adversarial use-- adversaria
l examples where you
have a prompt and then something is generated, you have that new
data set, what we're trying to do is get the companies and everybody to
kind of give us access to the model. And in exchange we say, hey,
we can build these data sets and we can give you these data
sets back into the ecosystem so you can lift the baseline of
safe AI for all of your models, because every model when
it's going up should have some minimum sort of thresholds
that it needs so it just works. How well
it works at the top,
that's a different story. But to Latanya's point, I think
that base has to be lifted. And I think the way we've
got to shift to big tech and so forth is to think, OK,
everybody should be able to lift it, which means, how can
we all work together to get researchers and academics
and so forth to help us do that? So that's something we're all kind of
collectively working on as a community. That's just one example of
how to address that issue. BHARAT N. ANAND: Very interesting.
We're near the end of the-- by the way, thank you all. This was a fascinating discussion. Let me just show the
other two poll questions and where the model response was. This was a question
about, have you started to think about the
implications of ChatGPT and generative AI models for your
own career/role/college major? So most are thinking about it. By the way, I don't know if you
have any suggestions for how undergrads should be thinking about what
to major or their own career choices. And I'
ll come back to
that in 30 seconds. And then the last question was, do you
believe the widespread use of language models like ChatGPT will lead to a
loss of human creativity in fields such as what we were talking about--
writing, journalism, content marketing-- or will they enhance our ability
to generate and express new ideas? So some optimism here. Let me maybe give you each the
floor for 15, 30 seconds each. What is the one thought,
question, or advice you want to share with folks as we start
to engage with these kinds of models going forward? Dustin, let me start with you. DUSTIN TINGLEY: Yeah. I think we in higher education need
to think about new types of curricula. So Latanya Sweeney started a track
within the Government Department, Political Science Department
on technology science that immerses in these issues. Here in our office, we've launched
a series on data and digital. And it's not about training
people how to code. That's great. We should be doing that sort of stuff. Bu
t critical thinking
around these technologies. Latanya happens to be featured in
the class that I taught, right? But the curricular reform can
complement the pedagogical reform that Rebecca was talking about. And I think that's a huge opportunity. The other piece real quickly,
we haven't talked about this. What are the business models that are
going to sustain these new operations? This takes enormous energy,
enormous amounts of servers to work. And I don't think a lot of them
really have a grea
t idea yet, and the advances that will be made
in being able to support the business models will be something to look at
that we haven't talked about today, but it's a fascinating topic. BHARAT N. ANAND: Rebecca? One thought, question, piece of advice. REBECCA NANCY NESSON: I would like
to see this technology revolution be different than the other
ones in terms of it helping us to address some of the
biggest challenges facing humanity. And I think it has the capacity
to potentially be deployed t
hat way, to help us overcome-- manage climate adaptation or any one
of these big challenges facing us. So that's what I would
encourage people to think about, is how to put it to work
to do something important. BHARAT N. ANAND: Great. Thank you. Latanya? LATANYA SWEENEY: I think the
question is, how do we maintain our control of our lives, our society? All of these changes by
technology are changing the way we live and work
and play, but they're also making the rules that we live by. And the que
stion is, we no longer-- it's not rule makers
increasingly or the laws that we have that determine our rules,
but the design of these technologies. And how do we restore
sort of control over that? BHARAT N. ANAND: Thank you. Vijay? VIJAY JANAPA REDDI: One
thing I'd focused on, it's like across the entire
planet we teach everybody doing any sort of computer science
or engineering-related thing how to program code. And we have heavy debates about,
should we use high-level languages, low-level lang
uage,
and blah, blah, blah. I'd say the course, for instance,
I'm actually trying to put together and scratching my head is, how do we
teach our next generation of students how to program with data? So imagine this course
on data programming, because these systems are all going to
be "coded," quote, unquote, with data. And today, we have absolutely no tools. We don't have a framework. We don't have a syllabus. We don't have a criteria for
how to systematically teach people to be able to build th
is. At the end of the day, there are
also engineers and scientists building these things, and the building
is really going to be with the data. They're going to code it with the data. So how do we systematically
think about that? That's one thing that really
keeps me up at night around this. BHARAT N. ANAND: Well, let
me just say thank you all. This was incredibly thoughtful and such
a delight to just hear from all four of you. I'll close with what seems to be
a narrative that is almost sort of
captured through this discussion. I think there's three sets of questions. One, what can these
technologies do today, right? Even just getting our
heads around this is hard. There's a second-level question of,
what does it mean for our enterprise, for our work, for our
jobs, for our roles, for what we do day to day, how we do it? But I think as all of you are
saying, it sort of leads ultimately into the third question,
which is, what does it mean for us as individuals, as humans? What does it me
an to be human,
as Michael Sandel, our colleague, often reminds us? Where does human agency come in? And those are hard questions. There's obviously some
issues that we haven't been able to talk
about, like Dustin, you were saying the business
model implications, which have huge, huge consequences,
as Latanya was alluding to in the social media context. Most of those outcomes,
I think, obviously were a reflection of
business model choices. Perhaps that's a
discussion for next time. Thank you all
. DUSTIN TINGLEY: Thanks, everyone. CASSIE MCGRATH: Thank you all. And thanks to all of you
who have joined us today. We hope you'll join us for
more Signature Events soon. Thank you for joining.
Comments