Subscribe
Share
Share
Embed
For Episode 2 of The Deep View: Conversations, we flew to Abu Dhabi to sit down with President Eric Xing of the Mohamed bin Zayed University of Artificial Intelligence, the world’s first AI-only university. We talk about everything from his unconventional journey to MBZUAI, misconceptions about the technology, the reality of progress in the field and the idea that, like the printing press, AI might usher in a new age for humanity, the age of empowerment.
SPONSOR:
Google Cloud: For all your cloud needs.
Go to https://cloud.google.com/startup/apply?utm_source=cloud_sfdc&utm_medium=email&utm_campaign=FY21-Q1-global-demandgen-website-cs-startup_program_mc&utm_content=the-deep-view&utm_term=-
EPISODE LINKS:
MBZUAI appoints Eric Xing: https://mbzuai.ac.ae/study/faculty/professor-eric-xing/
Eric Xing addresses the class of 2024: https://mbzuai.ac.ae/news/presidents-address-to-the-class-of-2024/
The printing press and the age of enlightenment: https://www.history.com/news/printing-press-renaissance
OUTLINE:
00:47 Introduction to MBZUAI
02:49 The Unexpected Offer
05:19 Making the Decision
10:01 A New Cultural Moment
13:28 Starting at MBZUAI
14:35 The Age of AI Empowerment
21:44 AI’s Potential and Risks
29:19 Regulating AI
35:40 The Promise of AI
43:39 AI as the Modern Microscope
45:19 Limitations and Misconceptions of AI
47:57 The Role of Language in AI
50:26 AI Literacy and Public Perception
01:01:00 Open Source vs. Closed Source Debate
01:15:01 The Future of AI Education
01:22:31 Conclusion and Final Thoughts
SOCIAL LINKS:
- X: https://twitter.com/thedeepview
- Instagram: https://www.instagram.com/thedeepview.co/
- LinkedIn: https://www.linkedin.com/company/the-deep-view-ai/
Creators & Guests
What is The Deep View: Conversations?
Artificial intelligence is a complicated topic, bound by many complex threads — technical science, ethics, safety, regulation, neuroscience, psychology, philosophy, investment, and — above all — humanity. On The Deep View: Conversations, Ian Krietzberg, host and Editor in Chief at The Deep View, breaks it all down, cutting through the hype to make clear what's important and why you should care.
Welcome back to the DeepView Conversations. Today for our
second episode, I traveled all the way from
New Jersey to the United Arab Emirates, specifically
Abu Dhabi, which is the home of the Mohamed bin Zayed University of
Artificial Intelligence. This is the world's first and only
university that is focused exclusively on AI research. And
right now I'm speaking with Dr. Eric Jing, who is the
president of MBZ UAI. Eric, thanks so
Well, thank you for traveling all the way here. I appreciate your diligence
and I love to chat
So I want to start with you, right? And as we've kind
of talked about before, the cameras started rolling, right? Your background
is in research, and you've been a professor, and you still kind
of wear that research hat. How do you go from a professor
of computer science to the president of MBZUAI
Well, to be honest, I never thought I'd become a
president of a university at any time. I'm
still right now affiliated with Carnegie Mellon. I've been working
as a professor in Carnegie Mellon since 2004. I
advise students, I teach classes, I
write grant proposals, and write
papers. In fact, I'm still doing all this right now, except
for classroom teaching. So research is very, very strong, very,
very important for me. That's why I haven't
thought about becoming a university administrator. During my career, when
I became senior, I did receive numerous invitations
and requests to be department chairs, deans,
even in one case another president, in
different institutions in the United States or in
UK and other places in the world. And I
didn't leave Carnegie Mellon. So
when I got an invitation for interviewing
here, I thought it's maybe
a learning experience for me to see the
place and also to maybe offer my my
insights or maybe experiences or opinions to
the leadership in here. During the interview, my interaction
with the host here is mainly based
on my identity as a researcher, as
a developer, and as a scientist. So I give them
basically the viewpoint you
know, from that angle, you know, how a university should be
built, should be operated, and what's
the best function and reputation
for university. And I also know that I'm
not the only one they interviewed, you know, they interviewed quite a few people. All
of them are much more senior and and
recognized that me. So I wasn't taking it too seriously
to be a real job interview. It's really an exchange, an
intellectual exchange. So I went back home. But then After
some time, to my surprise, they made an offer to
me. I wasn't ready at
all to take this offer because at
that time I was still very busy working with my students
and working on a few exciting projects. That was
at the beginning of the COVID. In fact, the
trip to UAE for the interview was the last trip I made
before travel shut down. in the States and
around the world. As we entered the
COVID season, things changed a
lot, even back home in the States. You
are locking down at home. You don't really get to
see your students so often. You teach through a camera, which
all were very new to all of us. And you started
getting yourself isolated from the rest of the world. Which
is good and bad. I didn't like that experience. I feel research
is important for science to be done collaboratively
with colleagues, with students. So I missed that part.
But on the other hand, it gave me the time to think. and
to really ask myself again where
I'm heading. So with this offer, which is
definitely non-trivial, it is supposed to
be bringing me to a place that I've never had previous
interaction and experience with, and also it's a job I've not
done before. So I consulted
some of my colleagues, mentors, and
friends. Professor Sir Michael Bradley, the former
acting president of the university, he
made me a very, very interesting case. He said, look, this
is more than just a university administrator. It
is about creating something from
scratch with a clean slate. And also, it
is in the region where no
such university ever existed. And the
country and the region needs such an institution to
really be the beacon of a new culture.
In a sense, I have the opportunity to to
influence and to shape a new type of culture, to
really help, you know, a
whole people, a whole country, you know, to
seek new directions, you know, or create new
results and also to redefine you
know, some of the decision process and the operation and
even culture, you know, in the region. So in a sense, there is
a chance for you to impact a country to
reshape the culture. So that's quite exciting to
me. I'm always surrounded by adventures in trying new things, and
this is definitely very new. Then I asked another friend
and mentor, Professor Michael Jordan, who has
been my advisor back in UC
Berkeley when I was doing my PhD. But then we kept the
friendship all the way. And he
also strongly suggested
to me to consider this opportunity more seriously. He
said, Hui, you are a professor
and a researcher with a good track
record of being productive and being creative on the
scientific part. But what could be
the next big stage? And that's what he asked,
right? Yeah, he understands that I want to
continue to do science, but bigger science also requires bigger
resources. And also, from a pure
career development standpoint, getting such
an offer at a relatively young age really actually
means that you have the
opportunity to experiment and also to try
different new ideas. In a sense, you have a chance to fail and
do it again, do it differently and so forth. So
he said this is a once in a life opportunity because
not every country are trying to build a university of
this and also not any time such
thing can happen. This is maybe just a once-in-a-life opportunity.
So these are the two strongest positive inputs
I can remember. There are other inputs as well that
are as positive. But of course, I also received a lot of warnings
and negative feedbacks. It is very
risky. It could ruin my career and reputation if I screw the job. or
I may run into a very different culture that I
just cannot understand and work with,
all these pragmatic issues and other things. So
I took a step back and I
took advantage of being in the height of the COVID with
very few people to talk to if I don't
actively seek out and this summer, in that thinking.
And in the end, I think I
will be living in my house probably
for the next God knows how many months and
years. I wouldn't be able to achieve too many
exciting things anyway. And this
particular opportunity is indeed once in a life, right?
And from a technical standpoint, I
started to see AI as a rising
sun on the horizon. the established
university infrastructure is already having
difficulties in coping with the new needs,
the new culture, and
the new evaluation and merit
metrics in this new discipline. But it is
very hard to change an established organization. So
even from a self-professional growth
and development point of view, it
is an opportunity to take your fate into your own hands and
also maybe create an environment that can help many of
our colleagues, students, and of
course the country and the people in
this region to a brand new set
up of such an institution. So I see that as a,
I start to appreciate that's becoming a huge opportunity that I can really
make a difference because I am an active researcher. and
it is not so common for an active
researcher and teacher to jump
into a university professor role. Usually, the
typical presidents you see are already somewhat
removed from the daily scientific life and
teaching life. Therefore, it
can be difficult for them to appreciate, you know, from
a close distance, the difficulties, the
needs, and all that, you know, in programming
and in organizing the university, right? So
all this converged to, you know, a
decision that, you know, I'm going to give it a try. And
then, you know, it took me a
year to make that decision. I actually felt
really appreciative that the university was
able to wait for me for that year. In fact, also the
Board of Trustees, they had the patience and also the trust
on me, you know, to onboard myself. I
remember it took me some effort, took
everybody a lot of effort, in fact, to even make the trip to here because all
the flights were unpredictable. And even getting
a COVID test certificate is also unpredictable.
But you need that to board a plane, which is a 24-hour fresh kind of certificate. a
lot of logistic challenges. So eventually I made
it to Abu Dhabi one day before the first
class of the students opened. I remember I got
myself prepared in a temporary
residence that I basically borrowed a
room from a very kind host, and
I delivered the opening speech of
that class, again, from a Zoom
meeting. Very
simple, not as fancy as today, the recording devices, which is coming
from my computer camera. So that's basically how the whole thing started.
At that time, the university was really, really very young and weak.
It has barely, I think, less
than a dozen faculty and only one
class of students, which was
very good. But at least I didn't have
any control or
knowledge about that because I was not admitted under
my kind of direction. So we literally walk
into the unknown and that basically
If you run a startup, you're going to want to pay attention to this. Startups, especially
AI startups, need good cloud providers. Lucky for you,
Google Cloud exists, and they offer an experience, the Google for Startups
Cloud Program that's tailored specifically to you. The program
offers eligible startups up to $200,000 in Google Cloud
credits for up to two years. For AI startups, that number is $350,000. Plus,
you get the mentorship and experience of Google Cloud's top-notch team. What
are you waiting for? Start building today with Google Cloud at cloud.google.com slash
startup. That's C-L-O-U-D dot G-O-O-G-L-E dot
Yeah, it's quite a journey. Now, about AI specifically,
right? And I know you've had an interesting journey with that. You started studying
molecular biology, and you kind of jumped into machine learning. You've
said of AI that it could usher in an age of empowerment. The
printing press ushered in the Age of Enlightenment, and
you've kind of equated the two as somewhat similar in how society-changing
AI could be. What does that Age of Empowerment look like? And
from where we're sitting now, what's the gap to get there? And
I'm glad that you asked this question because, you know, many people would ask
how to achieve artificial superintelligence and so on.
And then what is that, right? So, yeah, I think I
like to see any innovation and breakthrough from
a utility standpoint, what
function it delivers. and how such functionality
change people's life, change society, and
maybe even change civilization. So
it is in that context I call AI
to be equivalent to the
breakthrough leading to the age of enlightenment,
which is the printing press. There are many, many other inventions. The
invention of the steam engine, for example, discovery of
electricity, all these are huge. But I
think AI and
also printing press, all the invention of letters
and writing are at a different level compared
to all these technical innovations. Because if
you look at, if you put AI
and the printing press and maybe the
invention of paper and writing next to each other, they have something in
common. It is actually the backbone of human
civilization in their experience with
information. It's about how we
process and store and use
information in a
massively different scope of
magnitude. have
the technology of writing, use letters and characters
and papers. then
the information can be recorded rather than through
verbal communications. I
remember, is that Plato?
I think Plato is a student of Socrates. Socrates
didn't actually believe that knowledge, one should write
any book. He believed in in-person experience of
teaching. But it is
Plato who convinced him, or maybe who took
on the job of taking down all Socrates'
teaching and ideas into writing. And now we
appreciate how important that is, because that knowledge can
pass down to generations as it is. So in
a sense, it is a embodiment of
the antiquity time, information technology. But
at that time, you need to copy the book. And the
people who own the book need to be
literate and also need to have that kind of status, have
the opportunity to access to that. So knowledge belongs to a very few
people. Printing press is
a major advancement unlocking the knowledge from a few
people to everyone because it allows every
people to have their own copy of the book which began
from the Bible and they can form their own interpretations and
understanding and then after that people start to
think basically based on that from those knowledges and
then they write more books right so we have really
even the university evolving from a theoretical school to you
know study every topics Now, why
AI is directly related
to that train of evolution? If
you look at large language model, for example, it is trained not
on one book, It is not trained on
one sector of knowledge. It is literally trained on
everything ever written into language through
civilization and also in any media. In
a sense, you can imagine it is about turning all the
libraries in the world into one
device and then make it available next to your
fingertip. You can basically now easily get knowledge and
learn from it using this interface, which
makes the knowledge broadcast
and storage and accessibility to
a different level. Then AI also not only retrieves
knowledge for you, it actually started to be
able to use the knowledge to solve problems, as we can see. They can
at least answer many of your questions. And they can also
start producing, for example, code
and softwares. And nowadays, people
are trying to develop a lot
of applications in different sectors based on that, because
knowledge can be actionated to solve problems. And
in that sense, I call it an empowerment after
enlightenment. It is not about just giving you the
space to think. It is now giving you the tools to
make you a more powerful enabler
or actor in whatever occupation that you do.
And then, in a sense, leave you out space to
do maybe more challenging, intellectually
challenging things. And so that empowerment comes
together with even a next level of enlightenment
where, you know, everybody can do
very often time, you know, a very advanced level of
work without necessarily spending 20 years of
training, you know, in school. And then people should
not start rethinking, you know, what I'm going to learn now, you know,
in the in the next five years and what kind of problems I
should tackle. So we can refocus on problems
and maybe needs rather than just
focusing on repeating, executing
The difference to me between the large language models that
we're kind of dealing with today, right, and something like the printing press is,
you know, I think of all these technological innovations as levels of
automation, right? The printing press, as you said, got, instead
of having to handwrite books and pass them out, right, we could, we
have copies of stuff. it created more accessibility. The
jump to language models and
the kind of generative AI we're seeing now, to me, what you're starting
to do is you're starting to take the human out of the loop to a degree. Do
you see some sort of risk in that environment where maybe
for some people it might unlock them to explore greater problems? Is
there a concern that for other people it just acts as a crutch for human
intelligence, human creativity, because you have everything accessible,
I don't deny that there is a risk. In fact, any new technology you
know, comes together with both
opportunities and risk. Just like, you know, even
we talk about something that we are very familiar with,
you know, medicine, for example, you know, there are always a side effect that
we need to control. That's why FDA is formed. Genetic engineering, you
know, we use them now to dramatically raise the productivity of
crops. And we can engineer, you
know, into the crops, you know, or the
cattle's or the livestock's some nutritional deficiencies
and so forth. But you can also use
that technology to create very harmful viruses
or other forms of toxicities. So
it's really about how to mitigate and
how to regulate those risks. I think AI
is not an exception. But on the other hand,
also not a unique
threat that we didn't deal with before. In
the past, people figured out that there is a clear distinction between
the science, the technology, and the product, right?
And where you set the brick and
where you regulate is well
understood, that you probably should focus more
on regulating the users, on regulating the distributions, but
on the intellectual part, which is the science, and
maybe even the technology, which is the production part, you use
a different type of management. Typically,
for the science part, you don't even manage, or you'd manage very little, because
the creativity, once you shut it down, the cost
is too big to bear. In fact, many of the assumptions
about the current rush into AI
regulation discussion and even panic is
also based on an unclear understanding
about really the power of AI and how powerful
and how potentially dangerous it can be. It
is not well understood, especially by the general public.
Sometimes, even from a tactical standpoint, even
if there is a risk, how material that
risk is, and it actually had a big impact
on the risk and benefit calculations. So
I want to say a few words about really where we are in AI in
terms of its danger and its risk. Yeah, AI
is more than automation for sure, because AI can
automatically learn skills from training
materials. So it is a one-step beyond
just a programmed automation.
It does have the ability to acquire knowledge and to
become more and more powerful once
you plug them into some data sources and
also run an algorithm. But on the other hand, AI
also has a lot of limitations. It
is not a living creature that has self-identity,
self-drive, agency, free
will, and so forth. In fact, it is unclear to me at least where
all those will come from based on the current
architecture and our current mathematics of doing things. I
think down the road, new architectures beyond transformers, new
form of data beyond text and image will be feeding into
the next generation systems to create more
functions. But these functions actually are still quite
a few steps away from the type of higher
level intelligence or human-like intelligence I
just mentioned, like an agency, like identity, like a will, and
so forth. For example, how to even do more
complex reasoning, multiple steps reasoning. You want to
plan a shopping trip, you know, a travel, and
purchase a flight ticket, you know,
a plan on hotels. That requires multiple steps of
reasoning. I don't think right now the current
version of GPTs is able to do well on that job. Now,
talking maybe a slightly more challenging task. How to run
a store? How to run a company? How
to win a campaign, a battle?
These are reasonings which are typically a
hundred steps. conditioning on real-world
environments, and also conditioning on
embodied experiences from sensory and from anything. And
these are very difficult. In fact, you can probably already find
out that you cannot ask, for example, a GPT or AI
assistant to teach you how to swim, to teach you how to play the
music. Because why even different
music mentors, why
we have master classes? You know, those musicians, when they teach you,
it's not about reading from a book and say, hey, you apply
two grams of force onto a key and hold for
this many milliseconds and to produce a sound. It's
a very different experience of music teaching. And
I don't think right now our AI system is
even designed to do that. And certainly they cannot do that. So
I think the current AI system is still
on the steep learning curve to gain more capability, whether
it has the risk that people
are now, you know, fear
about. I think it's premature to say
that. Now people say disinformation, deepfake,
and all that, these are risks
not coming from AI technology. It is coming from
the users of the technology. It's just like a
gun. You definitely can cause a lot of harm by using
the gun. But is that the mistake of the gun?
People are still making better guns these days, and
for whatever specific purposes. But how
to regulate its usage is the key. So I
would like to remind policymakers and the
general public to really be
a little bit more concise and
careful about where the focus is
in such AI risk and AI regulation conversations.
The AI regulation conversation is very broad, right?
There are some instances that are trying to be very focused, such
as with deepfakes and job loss related to
deepfakes and the creative industries. There are others that
are trying to mitigate and prepare for these hypothetical risks
that we don't really know if they'll happen at all. And
And I want to add one more thing. In fact, even if you want to regulate
and counter that, you actually need AI to do it. Because like
the deepfake images and
the videos, the misinformation, it's just
too much for people or for a
manual devices or approach to discover
and to contain all of them. Unless you shut down the
computer, every computer in the world, they will be generated. And
As you just said, the problem with this technology and
regulating this technology is it requires
a global effort. If one country regulates really
hard, that's not going to stop China or other countries from
not regulating at all. And I think there's some concern about that.
the way you kind of framed the regulation question was to focus
Exactly. Yeah. That's why I think, for example, you
said about adversarial scenarios at certain countries
like China, will be doing what they need to do anyway,
and you cannot stop them. I think the best thing
that one can do is to advance
your own technology, in fact, to make AI even better.
Even we are under a material risk, we have
the means to take care of that, to counter that.
And do you think it would be an achievable thing? There
has been talk about the idea of international government, international
governance of AI that kind of supersedes any
individual efforts. You know, in the U.S. it's even more fragmented because
you have states doing things and then the federal government is trying to figure stuff
out. the EU is moving, and that'll impact where
companies deploy. But the idea of getting
every developer around the world from universities to
corporations to agree to a set of guidelines that maybe
Should it be feasible? I'm actually optimistic. It's absolutely
achievable. But on the other hand, I have to also
point out that if you look at the age of Enlightenment after
the printing press, it took roughly 100 years. before
all the warring states to eventually align
on some common principles and also to learn
how to live with this technology. In
West Europe, different religious sectors
fought to death just because a
different interpretation of a Bible principle. you
know, after all the people learn about it and have their own interpretations. But
eventually, you know, they learn how to settle things. And
also they learn how to live with a new society that
all can read and have many books and so forth. And
I can also name some counter examples where printing press
was actually banned. And after some hundred years,
those regions become much, much less developed. than
the West European countries. Why I mention this?
because I want to emphasize the importance of experimenting
and also the time. We are just barely five
years or maybe even three years into the age of
large-language models, right? So the AI era
has just started, and it is also evolving very fast.
We actually don't even see, you know, where it will platoon and
how a steady state would look like. And that's why I
think at this point, rush
into a premature set
of very stiff laws or regulations isn't
going to be very helpful because it may cause
missed opportunities and also maybe
an awkward misfit of all these
policies. with
respect to the real world, how they evolve. So I would say that
first, from a scientific standpoint,
we need to collect enough data points and
make careful observations and analysis. about
the opportunity and the risk, and
also the trajectory of the technological development. Thinking
ahead of the curve and be able to predict where
that technology is evolving, where I should put a
gate or a checkpoint down the road, rather
than passively you know, chase, you
know, a running kind of objects
and interfere with some of the, you know,
very organic and natural and creative and productive kind
of activities of scientific
advancements. The trick is just trying to figure out where it's going to go so
you can get ahead of it. Yeah, and it will take some time. But of
course, I'm optimistic that it wouldn't take 100 years. It wouldn't be as terrible
and as costly as the religious wars
Now, you talked a minute ago about we don't want to miss the
opportunity, right? And we've talked
about how impactful this technology can be. I've
been obsessed with this idea of trying to narrow down on
what the true promise of AI is, because like you said, I think a lot
of people misunderstand what the technology is itself. Part
of that is the fault of the kind of hype of corporations that
we've been seeing and social media is not helping. Part
of that is the fact that for a lot of subject
areas within AI, we don't have unified definitions, or people say
one thing when they're talking about something else. But
trying to narrow in on the idea of what is the promise of this technology and
how it can help people, because I think in
a lot of cases, people are afraid. People
are afraid about losing their jobs. People are afraid about existential
risk scenarios, which get kind of thrust
upon them from some of the companies developing this
technology. And it's set up as something that seems
to have so many, it's set up as something that's
dual use, but it seems to have so many negative risks that
I think for a lot of people, there is a question of why
develop this at all, if the risks can range across
this realm of severity. So to you,
Yeah, that has a close connection to really the
level of functionality that the AI will
be achieving step by step. And at
every functionality, you can accordingly kind
of define the cost and also
the opportunity. For example, Henry
Ford used to say that every American need to own a car. And
I think it's very bad news to the train
drivers and maybe to the carriage drivers. But
look, if every American needs to own a car, the number of
workers to produce those cars will be humongous. In fact, it
ends up creating a job. Now imagine that in our household,
we want to have a robotic maid to just take
care of all the household
businesses. Wouldn't that be an amazing product
that everybody wouldn't object to? I don't think right now people object to
washing machines and so on. So that's a
huge market. And the key technology to
get there is AI. You need to basically have a
robot that is able to plan, able to act,
be able to understand, communicate with you, and also
get things done. So we are actually, honestly, many
years away even from achieving that. But in achieving
that, I would imagine that everybody wants to be in
that business and join the production line and
make a lot of money out of it. Yeah, you will be losing, you know, of course,
you know, the clean ladies may be losing their job, but they will get a
better job elsewhere, because you will be joining the
production line in different steps of producing that amazing machine.
And also, you know, we talk about AI being
replacing some even white collar fancy jobs, like
a accountant, like a legal
worker. because we see HHBT are
passing the bar examinations, passing the medical examinations. Wouldn't
it be nice that it freed up these very smart people to
think about new practices in
those domains? So I think Sam
Altman, from his business standpoint, defined a
few tiers of AI. I
don't want to comment on whether those steps are too ambitious or
maybe less realistic. But
I think the way of thinking is very clear. It is aiming
for higher and higher functionality. And what
I can say from a technological standpoint, It
also will be dependent on
multiple generations of data, architecture,
and algorithmic evolutions. We're not there.
In fact, we don't have the tool right now to achieve even his second and third goals.
And within these second, third, and fifth goals, in fact, I can define
multiple concrete reasoning and
also acting tasks that requires
a lot of innovations. So at this university, in
fact, we were involved in some of this work, such as developing the
post-LLM models, including the word models,
that can even do a trivial thing of
simulating the word. If you cannot simulate the word, how
can you even teach a guy to pour a glass of water? That's actually not
written in the book. There are many other things that require
you to be able to simulate real-world experiences, and also generating
the next word based on your action on the real world, and
so on and so forth. And only in that way, you
can start to do thought experiments. You can do a new form of reasoning by
simulation, which could overcome some of
the limitations of the current language models. Then once you
have the word model, you can go one step further into
an agent model, where he has the word model as the brain,
but also he needs to have other mental devices, such
as perception, and
goal, and belief, and all these sort
of things. Each of them requires another
model architectures. So I can go on and speak maybe
for an hour about the technical steps that is required, even
to achieve what we probably can see as a very normal
and tedious level of intelligence. Then
I want to also say a few words about the other opportunities that is beyond
our daily life. which is the AI for science. Doing
scientific research is always a function of a technology environment.
Science requires more than brain, it requires the tool. Microscope,
telescope actually play the bigger role in advancing biology and
physics. AI is actually one such tool. Right now
we have, for example, in the space of medicine and
biology, there's a huge amount of genomic data, proteomic
data, and all sorts of orbit data that
is not meant even for people to analyze because they are too big and
too complex. Before the modern AI
of large language models and foundation models, we
can do very little in understanding and utilizing the
data. And now there is a new
movement, in fact, in computational biology of building foundation
models to distill in an unprecedented
way, you know, new form of knowledges and also predictability and
the simulation or capability of biological
systems. Imagine that. What if you can now
use an AI model to make proposals in
a very, very short amount of time of hundreds, if not thousands, of
drug design that can be applied to cure a
particular disease. And then you don't have to even apply the
drug to a real biological system
like a cell or a model organism. You can actually
also do the same in a word model of
the biology, which simulates all biological
phenomenons and the responses up to perturbation. You
can pretty much do this whole loop of innovation and
design and experimentation on a computer.
And that actually is now becoming a very, very exciting
goal for many of my colleagues, including students and
the professors in this university. Likewise, you can
draw the same experiences or expand the same experience in physics, in
agriculture, in many other sciences. So I think AI will
become, again, the modern day new microscope, which
tells you how to, helps you how to understand the complex data,
just like the classical microscope did. which is also about helping
you to understand information that you cannot see too small or
So you were talking about limitations of the current architecture. And
something else you mentioned that I want to tie into this as well is
that you refer to AI as our
new age microscope in the ways it'll
advance science. and not
as a species, right? And I think a lot of the source
of where the fear is coming from and where the hype is coming from, and I
think those two are very interrelated, is this language around
AI. And some researchers have taken issue with the
term AI at all because of this. as
the creation of a new species, as a species to supplant the
human species, as the kind of next thing.
And we are creating this being thing, replicable,
whatever that is. And it's vague if you try to pin it down, but the
way you're talking about it, with the limitations and the
focus on it as a specific tool, It's
very different. And so, you know, one, I
want to start with what those limitations are. You know,
we talk about hallucination, we talk about biases and
these kinds of inherent issues to the architecture. What's also
the importance of the language and how
Whether AI is a species and so on, these kinds of
discussions rarely happen among scientists
and engineers. I call them the consequence
of people living in different dimensions. Imagine
that there is a tribe somewhere in the Amazon jungle that
has never seen an aircraft or
maybe a car. and then all of a sudden they see
a person landing from
an aircraft. It is not it's
going to be surprising that they take that as a new species. In
fact, if we look at the ancient mosaics
and the pictures and sculptures in Egypt
and in other ancient civilizations, you actually see
a lot of such depictions of a god-like
figure or magical species, which,
I don't know, may actually represent a machinery. That's very
possible. So there is a key
distinction between technical and a
scientific discussion from a maybe
romantic and humanity discussions. For
example, this very word of intelligence. Intelligence is
not a very scientific design. I regret actually our
discipline is called artificial intelligence. because the other
fields are called physics and biology. It's a noun. Intelligence
is a noun, but it is derived from an adjunct. It's,
you are intelligent, and so on. It's a very subjective thing. For
example, imagine that I have in my hand a
calculator, but I didn't tell you. In fact,
especially for people who didn't know the existence of calculator, he would be amazed how
fast and how well I can do not just the trivial
calculations, but also square roots and the integrations and everything. I
can be called very intelligent, but it is a calculator. Once
you are in the same dimension, you see how it is made and
how it is supposed to function, you would be less impressed
and nervous. For example, large language
model. We built it. In fact, we are building it here as
well in MBCI. We are among the few universities, if not the only one in the world,
who actually pre-train large language model of our own. It
is a model. It is a language model doing
one thing, which is the next world prediction. So
in that function, many of the utilities somehow
can be encapsulated and dispatched. For
example, answering all your questions in
a certain scope. Next World
Prediction is like a linguistic simulator of
simulating how a person would
react to a prompt or a question through language.
It is literally a definable and a
measurable function. But human beings may
respond in a certain way. In fact, even between different human
beings, we do different. Some people recite some text. They
need to have a script to read from it. Some other
people improvise this. Some people are half-half. So
the approach of achieving that to a
receiver is actually less relevant. They draw on
their experiences. Now you sit on the end of the language
model, you see all your difficult questions, your difficult math
being answered. It doesn't prove that the machine actually understands
it, you know, or really, really digests it, or do it
the way a human would do. They can do it in a very different
way, and that way, by now, happens to be next-world prediction.
So that's why among scientists, among engineers, we
tend to be less amazed by this technology. And
we knew it is going to, for example, I'm not surprised at all that in one
day, in February soon, that a large-language model could
get a gold medal in the International Math Olympiad. Because
that type of reasoning, that type of problem-solving is
doable through a natural prediction model. As long as
you have a big enough brain to remember zillions
of words. you know, a prefix,
the sentence, and then the same amount next to that. But
I also know that there are some questions that they cannot answer, even through a
very, very simple interface. I've tried that. I reserved that
question list. If you are interested, I can send it to you. But I tested even
the latest strawberry. They still cannot answer. But
to a general public, they need education, I think. Right now, the
AI literacy is very, very different from
how much we know about algebra. Some people are very poor at algebra, but
they know what that is. They wouldn't be surprised to
see a calculator. But on the other hand, AI
literacy is much worse right now. When people see something amazing,
they literally get scared. I think there is a need to enhance
the education of AI. It's not about just math and programming. There
are other dimensions regarding AI utility, AI
measurements. and business and products that
people should know. Then they can have a much, much,
hopefully, authentic and realistic view about AI, even
without knowing the mathematics and programming. You don't need to be a mathematician or
That's the fascinating thing to me about AI is that it
is possibly one of the most multidisciplinary technologies that
I can think of. You have the technical aspects,
the computer science and the math, but there's also a lot of philosophy, human
psychology, neuroscience, linguistics, the
list kind of goes on. in terms of what you're talking about and the
core of what language models are and what they do,
which is a point that, you know, I appreciate you for bringing up because it's not
talked about often enough. And this kind
of race to intelligence and what intelligence is and what human intelligence is.
And the fascinating thing to me is the kind of stark differences between
language and intelligence and how these two
are not necessarily related. And
in interacting with these chatbots, right, I feel
like we, as a species, communicate our intelligence, our
thought, through language. This is our medium. This
is what sets us apart. And when we see something talking
to us in this medium of how we
convey our intelligence, it's easy for us to assume that it
That extrapolation is very natural for humans to make.
But even so, I guess it is not difficult
for us to realize that even if we don't speak,
we still function and we can still communicate. You have gestures,
you have many sensories that
can sense visual, audio,
temperature, all sorts of things. And not
all of this actually were described in language. In fact, language can
not be used to describe everything. There are many, many reasoning that
is existing outside of language. So
I would really draw a difference
between the functionality of communicating and
functioning linguistically versus
being intelligent. Because intelligence is, again, a very,
very subjective experience. It's a user experience. It's not a measurable
I think the fascinating thing about what we're seeing with
generative AI is that- There is one more thing, which
is oftentimes we call the soul. you
know, the spirit of a creature. I
have to say that I am very limited in my knowledge in understanding that
part. I don't know what that exactly is. Can we assume
that just by becoming knowledgeable enough, I
automatically become soulful? I'm not sure.
I think there's a difference between a machine and a biological creature.
And that difference is actually at least not well understood at this point.
I think the older generation philosophers had time to
think about it. They actually debated
on what is the status of human
reasoning and human morality and
will in the definition
of human versus machineries. I
think that discussion and that debate and thinking
should continue or restart because now we
have a new technology, which is AI. At that time,
they had this discussion because of the
latest invention, which is the calculus, the the
orbiting planets and so forth, which was fascinating. It may give you the
feeling that, wow, I'm like a god. I can predict the position
of the planets. I can predict the ellipse and so forth. And
then people start to think, oh, what is knowledge? What
is reasoning? What is conscious? So forth. I
think now it's time to rethink about that. Because with AI, which we
actually can make in a mechanical way using a
very, very simple training kind
of strategy, say, next world prediction, how
can we redefine what is ration and what
is mind and so forth? But I would
caution people from
making the jump of directly connecting
a very, very capable and seemingly
smart artificial system to
a human-like creature which has the
ego and also maybe
even the malicious intent to come after us. I think
there are a lot of steps, even philosophical ones and cognitive
ones, between them. And drawing an automatic extrapolation
is not very convincing, at least to a
The point you brought up about the soul, I find that
fascinating. And that's the kind of these differences
between knowledge and having been trained on every human work
ever written. the different facets that
make up what human intelligence is. And it goes beyond
knowing things. The stuff that makes humans human
I had one interesting debate with a colleague. For example, we
train generative AI, such as the Zara, for producing movies. Now,
let's imagine that the system watch a lot of Animal Kingdom
movies, like animals fight each other, eat
each other, and so forth, so that they can produce similar movies. Along
the way, are they going to be learning the will of
killing and of predator and prey
type of mindset? Well, nobody actually
knows, at least at this point. But it's a fascinating question to
Intention is a big thing here that's missing.
But there's also a lot that's not known about language models.
The real risk I do see in a potentially
out of control AI system is not about
that they kill people and so forth. It's about their
environmental footprint. Imagine that if
you do have an AI system that is told to
infinitely iterate on some very,
very difficult objectives that actually can be measured, but
they can go very, very big. What if the machine decides
to keep training themselves and burn a
heck of energy and so forth, and causing really environmental
disasters? To me, that's a real risk we have to be careful
about. There has to be checkpoints and the bricks, so
that at least a person should be allowed to turn it off. If
you want to build a system that is never having a switch, probably
no good. But these are something you
can already bear in your mind when you do the design. And
we could call it also a regulation in the science space. But
I don't want to overgeneralize that. It's not like
everywhere. It's only a very, very
definable and measurable risk that is indeed very,
very grave. But on the other hand, I think
Right. And a lot of these risks come from putting
these systems in a place where they can make decisions without human oversight. And
I guess the immediate and obvious remedy is don't do
that. Maintain human oversight. You won't have those problems.
Yeah, yeah. That's true. So I
think it's an ongoing discussion. But again, there is still
a difference between the scientific kind
of brain work, or maybe even
the technological experimental work from
the production and deployments and
the usage. I would imagine
that it's more effective and productive to focus on
the latter. Also, one more thing. I
think AI, compared to other sciences,
has the advantage that it is still one step
away from the physical world. Imagine that you are a biologist, you
do virology and genetic
engineering. Well, the products of
your experimentation is a physical being. If
you somehow smash the glass, or you somehow cause
a lab leak, it actually does enter the
environment. And they will stay there, or they could spread,
just like what we see in COVID. So these are very mature
risks. And therefore, it requires rather
stringent regulations and control. AI maybe
conceptually have that kind of same risk, but
it is fortunately still in a digital
and a virtual environment. In the worst case, you just
throw off your iPhones and shut down your machines. You go back
to the cage, the caves. We will still
survive, right? And without, you know, for
example, computer. Of course, I know it's very difficult these days, but I
mean, it's a different level of risk. That
existential risk is different from a biological disaster,
a chemical skill, a nuclear kind of explosion,
Part of the regulation question, and all
these topics are so passionately debated in the public space.
is the idea of open source versus a lot of what
we see from the companies, which is super, super closed source.
We're not even going to tell you what we trained on. That has to come
out through reports from investigative journalists. And
it trickles out and is fought over in court because of copyright concerns.
But the debate hinges on this idea
that science needs to be open. Science
needs to be explainable and replicable. And if
it's not open, it's not really science. And it's the distinction between product
and science. But there's also concerns about
what open source could mean for security risks. And you
were talking about chemical and biological situations
and how AI open source can maybe help bad
actors in those realms. So as
it relates to that and making sure that nothing bad happens, what
Well, I don't
want to reiterate the philosophy in science
that there is a fundamental value of share
and open and so forth. Definitely, I'm for that. More
than that, I actually don't think secrecy is
a recipe for security. I
think there's a fundamental issue. People thought by keeping things secret, you
become safer. I doubt it. I
don't think history gave us too many examples of that. Because
eventually, you will lose the secrecy. I think having
more people to understand the technology, to be able
to investigate and experiment creates
a much larger pool of people
and also a large pool of knowledge to contain
the risk if that risk happens. That's kind of, again, maybe
a more philosophical and abstract argument. But
I can give you a few examples. Right
now, we don't actually know exactly how
GPTs are trained. Do
you think that makes the technology safer or less safe? Well,
I don't know. I feel less comfortable
if I'm using something that I have
no knowledge from. Yes, I'm not going to break my car and
do it myself or understand how
it is made. But at least I know that there
are more than one manufacturer who knows how to do it. If
my car breaks down, like I have to drive a Mercedes, I don't have to go back
to Mercedes. I can go to any car shop to have it fixed. That
makes me feel safe. So under the same principle, I think important
technologies, I'm not even arguing a commercial reason. I think many of
the arguments was because of a commercial interest and other
geopolitical interest. But from a security standpoint, open
source has the advantage of
sharing the knowledge to more
people, and also keeping the
technology transparent for scrutiny, and therefore, for
a better opportunity to regulate it. So I think that
part is very important, even from a pure security
standpoint. Now, of course, there are even more
arguments about the economical and the scientific kind of
value from it. Economy is obvious. You don't want
a lucrative and important technology to be
monopolized by one of your companies. That's why in
the US, you have the antitrust laws and the anti-monopoly laws,
right? So if your company, Rockefeller, once
they own all the petroleum production, the
government orders it to break it. There is always an interest
of healthy competition among entities for
key technology and key products to actually better the
quality of technology itself and also to make it affordable and
accessible to people. So that's a commercial argument I want to make. Open
source gives startups, even
universities, a share of their impact
and also their ability to
innovate in this market. That's scientifically even
more important. I think no science actually gets
developed very well if it's kept in secret. So
for the sake of advancing the science of artificial
intelligence, especially in the United States, definitely the
U.S. is right now the current leader. I
think by having the open
source community very robustly
grounded in the United States, actually end up helping the
United States to be the leader for
the past many decades in IT technology. Look at, for
example, the development of
operating systems like Linux, the search engines.
and many of the novel IT innovations
is because there is this openness of
technology to be played and experimented and
advanced, not just by a designated player, but
by the entire academic and
industrial community at large. I think that capability and
freedom is nowhere to see, in fact, elsewhere
in the world, not even in Europe. In
fact, that's part of the reason I become excited to take
on this job in here, because I do like to
see such a mindset and such a value,
a principle in education, in technology, to
impact regions which are yet to be impacted. So
here is a young country that
wants to have a voice and have a stake in the global technological
advancements. And they have the
resource and also the ambition and also the enthusiasm of
being a leader. And absolutely, I think by
you know, moving the,
maybe I'm lacking the right word, by learning or
maybe by advocating and advancing, amplifying in that
kind of openness, that kind of transparency, you
know, from the very beginning of the foundation of this university and
of the talent we are training, I believe is
at the end of the day helping, you know, to grow this technology. And
that's also why in our university, yeah, we indeed take a very, very
open, but of course, compliant and
regulated approach to develop our AI.
You may have heard about, of course, everybody heard about Lama from Meta,
right? And it's a huge benefit
and support to
the community. And they benefit a lot, a lot of researchers working
on that and maybe inspired
a lot of new advancements. In here, we
have a project called LLM360, which actually went
even multiple steps further. We not only open-sourced the
model weights, we open-sourced the
checkpoints and also the code that
you use to produce the model. And our goal is
to let researchers to be able to reproduce the result,
good and bad, because otherwise you can only use
it and augment it. But you
still have this dependency of waiting for the next version, you
know, for you to rebuild everything from scratch. And
in our case, We want to make the community even bigger
with this higher degree of transparency. And I
think, does that give us any risk? Well,
you can never rule out the possibility of a bad player somewhere
in the world to take that and turn it into somewhat
a malicious tool. I
don't think, you know, that's going
to be stopped if we just close source it. People will
figure out from else places. But on the other hand, by
opening the technology in such a way,
we end up having more people understanding how
to augment and how to manage
and contain the risk from this machine and how to make
the machine better. So again, it's a debatable topic, but
I just want to share with you my view from a developer and
a researcher who makes these kind of technologies, I think
And we talked about this a little bit, but the thing that I think is important
in the realm of AI literacy and helping ordinary
people who don't spend their time reading
research papers or writing research papers about AI, What
are the biggest misconceptions that you see people
make about artificial intelligence in general
Well, all the things that we just discussed, like taking
AI as some sort of a creature that
could spin out of control from human hands, is definitely
one misperception. And
also exaggerating the power
of AI in real human
life, and also in scientific exploration, In
my opinion, it's also a misperception because AI
is very powerful, but
it is powerful in certain dimensions. For example, it is very powerful now
in playing the game of chess and Go. But it
doesn't mean that it plays other things well, like planning
your trips. So it takes time to get there. Another
misperception is that AI is
only about large language models and the GPTs. In
fact, AI is a very, very broad technology that
is more than actually we
expected to be coexisting with
us already. Our search engine is also an AI tool. And
I don't think people want to ban such engine. And
also, now everybody is driving
a car with a GPS or
with your iPhone next to you to
show you the road. I can tell you that I'm old enough that I live
in an age where I use a paper map to drive. It's painful. I don't want
to go back that age. People don't realize that that function is also
from AI. So AI is already everywhere. So
using the GPT as the
main or the only representation of AI is
itself a misperception. And therefore, the discussion around
AI sometimes is
very, very confused and ambiguous. Maybe we
should say that, how about we discuss large language models? That's actually
a lot more precise. and
better scoped. But when you say AI, literally
you are talking about how about we take a Google
map from your iPhone. That's a part of AI. And
who knows what will come along that line, the next product. I think
self-driving is one of the major things That
is coming very close to what
it's already realized. It's just like right now the policy is
lagging behind for its deployment. In
fact, I think a policy is more needed over
there in terms of deployment. Because if you look at the statistics, the
accidents committed by an autonomous driving car actually is
lower than people. And the
reason why it is not widely adopted is a policy and
infrastructure issue, not necessarily a technological issue.
And so I think there are a lot of examples in
such a space where especially decision makers
and policy makers will be benefiting from
having scientists and developers next to
them to be part of the discussion. I have to say
that I've been to events in various organizations
in the government bodies where a scientist
like myself could amount to
5% of the participants, or even less. So
in that environment, it is amplifying, in
We talk about people kind of trying to educate themselves better. And
like you were saying, it's really difficult. A
lot of this is ambiguous. A lot of the language changes depending on who you
talk to. What should people
be doing about AI right now? How
should they be kind of going about thinking
about it? I think there's a
degree that some people are afraid that it's too complex, too
out of reach to have an opinion on it or to think
about it too much. What would you tell them they
Good question. I
have to say that I don't want
to push the blame to the public just because they don't do
enough. I think the general public, even
including decision makers, policy makers, I
genuinely believe that they wanted to know more, but they don't know where
to learn and how to learn. So our educational system
should be the starting point that our current curriculum regarding
AI or computer science probably it's time to revisit
them. In fact, for the past, I
would say, several decades, if you look at the curriculum,
there hasn't been a fundamental change. It's literally just about how
much I teach along the same line. Which program
language? Okay, I can start it from Fortran in the 50s,
and then C, and then Java, and now Python, and so forth. But
it's still about teaching a different language. mathematics, well
there are new material coming in, let's retire the older ones
and bring in new ones. But yeah, you need to know linear algebra,
you need to know calculus, and so forth. So it's still the
very, very conventional way of thinking, and the statistics, for example,
are still using a year to teach regressions and so forth. With
now AI and with functions like
GPT or search engine and so forth, in
fact, we already are in the crisis where some people are using
GPTs to write scientific reviews of the submitted
publications. And people are getting very frustrated,
but you cannot stop it. It's happening, right? So then your attitude towards
even published papers may need to be augmented. Are
they really the product of the the researchers or
it's a mix and so forth. So I think this awareness of
the changing technology should be first taught to the public,
you know, in some curriculum. And that material does not require you
to know calculus and know linear algebra and know all this fancy
mathematical stuff. And in fact, it could be taught even
in elementary school and high school. So I think bringing
the AI into the background of
all teaching materials makes them not a specialty,
but a very, very horizontal, fundamental
science is probably where we want to begin from. Start early
and start very, very simple but broad to
create awareness about what we
already are using AI to do, and how
well they can do, and what else we set
forth for them to do. That kind of high-level knowledge
can be already taught to the general public
without any emphasis on technical specialty.
Then you can talk about the connections of
AI with all disciplines, not just AI
for science or AI making a product. For
example, artists are already using generative
AI to create visual arts,
music, and many other things. How
do we appreciate and interpret these
kinds of outcomes? and how can we redefine copyrights
and maybe authorship
and all that kind of things. I think this kind
of issues are already happening, but
very few people really put time
to think about it, or not even maybe knowing it. So
I think the AI literacy is really about awareness
of such an entity, such a
capability. We all know, for example, the capability of linear
algebra, or not linear algebra, just algebra. That's why
we are not afraid of a calculator, because we actually know algebra, even
though we don't know how to do algebra, but we know what is algebra. Right
now, we don't have such a literacy in AI. And
that's where, from the educational standpoint, I
would like to see universities and even
going further earlier, high schools, elementary schools, start
thinking about that. My son is in high school right now and
I sometimes look into
his curriculum and try to understand how
much their teachers and principals actually know
about all this. I think it's fun to keep an eye on it. Then
given that's happening, I have more to say to
the general public. Hey, go read
such material. Go watch a few informative
videos on YouTube and other places. But on
the other hand, go ignore certain kind of information that
you get from all different sources. There is definitely a
difference of good and bad information. And there
is also a judgment that you have to grow to be
able to tell what is really
valuable and authentic versus
what may be something suspicious. I think, again,
I'm not a fan of just
cleaning up the things to make it like a
hospital and a clean room. In fact, if
you look at evolution, both in
terms of a biological entity or even evolution
of our brain, You evolve and
you become healthier because you have adversaries. You
have all these viruses, you have all these pathogens, you
have actually all these misinformations from
your world. And you co-evolve with it. You
end up developing your own immune to be stronger. And
keeping you away from all this isn't going to make you healthier. In
fact, it makes you weaker. I don't think the politicians
and the general public appreciate that. They want to be as
safe as possible, away from all risk and all
kind of pathogens, psychological pathogens, you
know, or biological pathogens. I think it's going to weaken
our civilization. And then when you have an adversary, we
talk about China, we talk about some other countries, what if they don't do that? You
will eventually run into an adversary that is healthier than
you, and that will be even worse. So I would really
advocate a culture of openness and
also of brevity to embrace this
reality and then try to take on it. Thanks so much,