Thursday, Oct 27, 2022 • 30min
12. How To Understand Where We Are Heading With AI, With Professor Marc Mezard
Play Episode
Today we’re focusing on artificial intelligence by talking with an expert who researches the forefront of technological capabilities.
This is for any listener who wants to better understand AI applications in business, where the field of AI is heading, and what breakthroughs may be right around the corner.
Here's our conversation with Marc Mezard, Professor of the Department of Computing Sciences at Bocconi University.
Looking for a new guide to drive innovation and change? The Talent Show is a new podcast series from FT Talent, a hub of innovation from the Financial Times. Hosted by under 30s for under 30s around the world. Each episode we have important conversations for you and with you. We speak to experts in different fields, and bring you in to ask them your burning questions and delve deep into the topics that really matter to the younger generation today, find inspiring tips, analyse trends and bridge generational gaps. And we didn't just rely on our own curiosity - we invite our audience of bright students and early career professionals from all over the world to ask questions directly to our guests. The FT Talent Challenge is a competition from the Financial Times that invites bright young talent from all over the world to pitch solutions aimed at solving our most pressing business challenges. This podcast gives you a taste of the creative, educational and entrepreneurial atmosphere at FT Talent Challenges. FT Talent is a commercial division of the Financial Times. This first season of The Talent Show Podcast is in partnership with Bocconi University, a leading university of business, economics and management teaching and research.
The FT Newsroom is not involved in its production. Our GDPR privacy policy was updated on August 8, 2022.
Hosted on Acast. See acast.com/privacy https://acast.com/privacy for more information.
Read more
Talking about
Speakers
(4)
Marc Mézard
Virginia Stagni
Jean
Show more
Transcript
Verified
Marc Mézard
00:01
The new developments of the last 10 years in
artificial intelligence
are starting to reshape the world the economics of the business etc.Share
Virginia Stagni
00:14
This is the Talent Show, a new podcast series from FT Talent, a hub of innovation from the
Financial Times
, hosted by under-thirties for the under thirties around the world. This first series is in partnership with the university, a leading higher education institution of business and managerial advancements high in Virginia Stagni, and this is the guide you need to drive innovation and change.Share
00:43
Today we're focusing on
artificial intelligence
by talking with an expert who researches the forefront of technology capabilities, visits for any listener who wants to better understand AI
applications in business where the field of AI
is heading and what breakthroughs maybe right around the corner.Share
00:60
Here is our conversation with
Marc Mezard
, professor of the Department Of Computing Sciences at Bocconi University
.Share
01:11
Thank you very much Mark for being with us today.
Share
Marc Mézard
01:14
Very nice to see you today and have this discussion.
Share
Virginia Stagni
01:17
Very nice to have you on the Talent Show, the topic that we would like to explore with you today is something that has been a bit of a buzzword every time that we are thinking about what are the new jobs the next generation is interested in.
Share
01:31
And definitely looking at
AI
from your point of view and the fact that in the business school in Milano
there is attention to AI
is such a pleasure to learn from your experience and from any tips you might give to our young listeners out there, that are interested in building a career in AI
.Share
Marc Mézard
01:48
I'm sure that there are many of them who are interested in that. We'll try to answer their questions.
Share
Virginia Stagni
01:53
I would love to know more about your personal journey. How did you get into this field?
Share
Marc Mézard
01:59
Oh, I got into this field because I am a theoretical physicist working on complex systems. So complex system is a collective behavior of systems which have many, many components. And in some sense you have a complex system when the total is more than the sum of the individuals.
Share
02:16
So, I was working actually, for instance, with a lot with
Giorgio Parisi
who got the Nobel prize in physics last year and who has written these books about storms, these flights of birds, how they coordinate how they move and how the collective motion is much more than just the sum of the individual flights.Share
02:36
So, I came from that, and then at some point already, starting in the 80s, we went into the collective coding of information, how information can be encoded in a collective system in which you have many single elements. How information emerges, how can this be treated?
Share
02:54
That was the very beginning of what was called
neural networks
. It has had period which was more silent but in the last 10 years or so, this is a big boom of artificial intelligence
is based precisely on this collective monitoring and treatment of information.Share
Virginia Stagni
03:11
How did you start your research?
Share
Marc Mézard
03:14
Oh, I can tell you, I started in theoretical physics trying to understand the properties of very strange materials which are glasses. Glass is a very strange material. You do not realize that because you have seen glasses everywhere.
Share
03:28
It especially in the sense that if you zoom on the glass you see the molecules that build the glass. But in general in solid the molecules they are very well ordered, they are well aligned and that's a crystal, or that's a metal in your glass. They are disordered like you would find in a liquid but they do not flow like in a liquid, they do not move around.
Share
03:51
So I tried to understand that I went to
Rome
to work with Giorgio Parisi
on these kind of things and we made big progress at some point we realized that the ideas that we were developing that really ideas that have to do with how you describe these molecules in a glass, each of them is different from all the other ones. Each of them sees a different environment. So it's like a very complicated society made of billions of small molecules, each different and each reacting differently.Share
04:22
And we realized at some point that it could be used to study some problems in a very different fields in biology, in computer science. And I decided to move my activity towards computer science and especially the branch of computer science which has to do with how information can be processed in an assembly of a lot of very simple elementary tools like artificial neurons and that was the beginning of
neural networks
.Share
04:51
And these
neural networks
, they have become now the big tool of artificial intelligence
. I found myself in this story very happy to be there.Share
Virginia Stagni
04:60
In terms of succeeding generally in
AI
, do you think that you need a Ph. D. today?Share
Marc Mézard
05:07
Well it depends what you mean by succeeding in AI. I think there are really several fronts of
AI
at the moment. It is obvious to everybody the new developments of the last 10 years in artificial intelligence
, they are starting to reshape the world, the economics, the business etcetera. And so there will be need for people who have some knowledge of what AI
can do. This will be needed at all levels, in all kinds of jobs.Share
05:36
And then there is another frontier in some sense which is, where do you develop the new tools with the research around
AI
? This is technical, it's mathematics, physics, statistics etcetera. And for this certainly, yes, you need a Ph. D., and you need more than a Ph. D., you need a few years of research.Share
Virginia Stagni
05:55
We had to give a bit of more theoretical definition of what we mean
by
AI. Would you mind explaining in very simple terms, what do we mean today for AI
?Share
Marc Mézard
06:07
Well I will focus on one aspect of
AI
which is really the one that has been expanding so much and which is triggering this revolution that we see in the last 10 years. This is called machine learning
. Machine learning
is you are given a certain task typically a cognitive task that is you see some images on the images you are asked what is there on this image? Is there in the simplest case when someone is there a cat or is there a dog?Share
06:34
People have tried to write by hand some computer programs in order to do that. You know, they were trying to see if there is some special feature that is a cat or a dog etcetera. And it never worked really well. And the new tool is
machine learning
.Share
06:49
Machine learning
is just you let the machine learn by itself that if you present a big database of images of cats with label cat, and images of dogs with the label dog. And you have a machine which has a lot of instructions but the instruction they are not in order, they are kind of random at the beginning.Share
07:10
And the machine gradually will make the instruction one by one fit in such a way that it does very well on all the database that you haven't given it. And then it becomes a smart machine that is able to identify cats and dogs. So that is the typical treatment of the
machine learning
.Share
07:26
It's a big revolution in the sense that you use machine, you do not give yourself all the instruction, you just give the instruction to the machine: how you can learn. And then the machine learns from a large database of examples. And this has been applied to image recognition, image segmentation, speech recognition, automatic translation. The number of applications is enormous; is growing every day.
Share
Virginia Stagni
07:51
What do we still don't understand about
AI
and about the AI
field, is still something that we need more research on and then maybe your mentoring and inspiring your students and your students to pursue?Share
Marc Mézard
08:06
Oh yes, this is a very interesting question. What we are seeing at the moment. What I have tried to describe briefly is technological revolution. It is already there. We can see its application, there are incredible tasks - I was talking about translation about identification of images, but identification of images means also the ability to have self driving cars or things like that - this we see all the technology which is really going forward.
Share
08:35
At the same time. The real understanding of what is happening in
AI
in modern AI
devices is still missing. So it means that when you have this machine that has learned to task by looking at a myriad of examples, and then it does well, you can measure how good it does on a new task and you find it does well.Share
08:58
But there is no explanation. The machine cannot tell you, "I think that this is a cat because this and that and that". The machine cannot tell you. "I have decided to translate this sentence from the French to the English in this way because of this context", does not know that because the machine does not have this information.
Share
09:17
And so this question of explain ability of the decision of the machine, I think it will be very important for the future of applications,'cause when you start to have application which have a direct impact on the public, the people need to understand on what basis the decision has been made.
Share
09:34
If you have a recommendation system that help company to hire some new employees, for instance, this already exists. If you feed in with the CV and it answers by ranking the CVs and saying "this is easy". That's something, but if there is no explanation behind this ranking, it's a very weak way of doing things, we need to have explanatory tools.
Share
09:58
And explanation is very far away at the moment. And because the theory is lagging behind, we have had a technological breakthrough, but the theory is lagging behind. So we need much more research in math, in theory on these devices.
Share
10:11
That is also very important in order to give some guarantee. Guarantee that it will not go crazy. If you have a self-driving car which is analyzing images based on some
AI
device for analyzing images, you would like the car will not do something completely crazy if it finds a situation that it has never encountered. I don't know... you're driving and there is an earthquake. Well the car has never seen that. I don't know how he will react, but certainly I do not want the car to do something completely crazier if this happens.Share
10:43
So, the fact of having a theoretical understanding it means being able to explain what is taking place in the machine, it means being able to give some guarantee that it will not go in a completely crazy situation.
Share
Virginia Stagni
10:58
It's about the critical thinking point, right? And it's how do you can input that into more advanced technology is something that it's quite interesting even from the decision-making process, but as well, like when we're looking, for example, at news and new tools that we can use in the newsroom in deciding what you should know every day.
Share
11:19
And the kind of selection and filtering of news is still something that we believe it should be and could be human-centered, because the technology doesn't do yet the job of a critical thinker but is normally part of the newsroom. And that's been a huge debate, for example, when it comes to how does it look like the newsroom of the future or a media group of the future.
Share
11:43
It is so interesting that when I joined the
Financial Times
, it was a year where we just started an experiment of having, you know, little AI
tools allow me to say. So that was trying to recognize images on our paper to try to understand it. We were balanced and equal in getting men and women and as well diversity in our pictures.Share
12:06
And it was interesting to see that sometimes the machine made a few mistakes. The
Financial Times
engineers tried to give a color as well of the on this picture. Is it a sad man or is a happy woman? Trying really to balance a bit the sentiment around genre and diversity. Something very tangible for me and pragmatic because it was giving recommendations to the newsroom on the images to use.Share
12:31
Have you seen something that made you excited about your field in terms of very pragmatic applications that maybe you would like to share with our Talent Show listener?
Share
Marc Mézard
12:42
Well what I would like to say is that at the moment there is something very important taking place, which is that
AI
really these days these months is becoming more and more important in science itself. It is having an impact on scientific discoveries.Share
12:58
For instance, there has been this program called Alpha Fold that was developed in the recent years.
Share
13:04
Alpha Fold is a computer program based on
AI
which is able to take a sequence of a protein and tell you what will be the shape of the protein. So the protein is a molecule, the molecule is made of some basic amino acids. This amino acid, they come in 20 letters. Okay, so you have 20 of them, you have a letter for each of them.Share
13:26
And so you have a text which is just a one line, a long line gives you the sequence of letters. And this text, we were not really able to read it Well, it required an enormous effort and now we have a tool in which you give the text and the output of the text will tell you "this is a molecule that is done like this and like that", and it will expose this site and this site combined to something else and it can have that action.
Share
13:51
It looks like nothing but has a tremendous impact on old biochemistry. And biochemistry, it means developing new drugs for medicine. I mean when you have a good pro tain that is able to bind to something that's exactly the process by which you develop a new drug. We are just at the moment where we are able to read this book. We read the book of proteins.
Share
14:14
So it's a very exciting moment. It's something that we didn't know, as if we had a language. Of course we knew the basic rules etcetera. But from these basic rules, we were not able by reading the book to know what it will be able to do and we are about to do that at the moment.
Share
14:31
It is like, we are learning a new language, is learning this new language will allow us to read the great book of life. So it is a fascinating moment and I picked that example in biochemistry because I think that it is a very big issue. It's scientifically, it's really an important progress. But when you can see the impact of
artificial intelligence
development in many other areas.Share
14:55
I was discussing very recently with colleagues who are in chemistry and they are trying to build new materials, porous material that will be able to absorb CO2. So if you absorb the carbon dioxide, you know, it's something that is very important for the future of the planet other which will be able to absorb hydrogen. Having a hydrogen which is stored on a solid in a very efficient way is also the key for having green cars that will work only with hydrogen.
Share
15:27
So there is a lot of science to be developed on which you can rely on a lot of data because there have been many studies but you know, the number of materials that could be synthesized and tested, they are enormous.
Share
15:41
So you cannot do that. Each time it's a very long work of chemistry to synthesize a piece of new material that nobody has seen. So, if you can predict with
artificial intelligence
, "Look synthesize this one, the one that has this composition", probably that will be one that will be a very good absorbent for hydrogen. So, it changes a lot of the science developments and the technological developments.Share
Virginia Stagni
16:05
We say at the
Financial Times
that the great catalyst in our change has been in the distribution side. And definitely data had an enormous, huge role, being able to understand our readers, to understand our customers made us do different editorial and business decisions and take certain directions.Share
16:28
We then say - and we actually read this quite a lot - that information is the new oil and capitalism. Do you agree with this? What are the questions that capitalists should start thinking when they are managing the data? And I think let's raise the ethical question. And how from your side, how can you influence ethical decision-making or at least stimulate the conversation around ethics and data?
Share
Marc Mézard
16:57
Okay, first, yes, information is the new oil. It is very nice statement. I think it is partly true, but I would like to compliment it with something. If you have more information, it means more data, it means more capacity to be able to do to address the customer with more precise with what the customer might want to get, etcetera etcetera. So you are more powerful when you have more data available, if you are able to treat the data.
Share
17:29
So the new
artificial intelligence
is able to treat large amount of data, kind of... it's able to treat it in the sense that it's able to answer relatively simple questions. Not very complicated questions.Share
17:44
I have one caveat that I want to emphasize, which is important also. I am a physicist and in physics there is something which is very important which is the energy, the energy is conserved. And so oil is energy as we know oil and gas, our energy that we use every day, we know it so well.
Share
18:03
And so information is wonderful. Information is much more linked to another aspect of thermodynamics which is called entropy. That's a long story. But the energy will not disappear. The energy problem will not disappear because we have more information. Not at all, we will always need energy.
Share
18:20
And in fact the thing which is very impressive is that we have seen in recent years all this progress of
machine learning
. They are based on what? They are based on larger database, more and more computation.Share
18:33
And the amount of computing power that is dedicating to
AI
is increasing very rapidly. The amount of energy of electricity that one uses in order to develop these new algorithms of machine learning
is increasing very fast also. and we will reach a saturation at some point because of that because of the energy it will come back, it will bounce back.Share
18:56
So certainly that is a direction of research that I like a lot which is parsimonious
machine learning
. How is one able to develop new machine learning
algorithm, new machine learning
devices, tools, ideas, that are not relying on the fact that you have 10 billion data but that you have much less data but you exploit them in a smarter way?Share
19:24
And so that is "parsimonious learning". I think it's a it's a very interesting research direction because we will hit this frontier of the size of database that at some point becomes unmanageable just from the point of view of energy consumption.
Share
Virginia Stagni
19:39
One question for you - and I hope this comes across in the right way - because I want to make this a bit relatable to people like me that try to code a few times, trying to learn how to do it and they failed.
Share
19:54
So for people that are a bit more developed on the right brain, how they can learn to at least understand the tools and get more hands-on in the
AI
field, what would you suggest them to do?Share
Marc Mézard
20:09
That's not an easy question. I'm not sure I have the answer to that. I am impressed that you try to code by yourself, I'm not sure this is really the only solution, although the exercise of coding in itself is rather interesting, because any example of people who try to code, in general the first try you make a mistake, it does not work, it does not produce what you expected to produce, at least that's what happens to me after decades and decades and decades of coding. So it's still true.
Share
20:41
So you learn at least something. You learn that there is a language there which is very precise and when it needs to be very well tuned.
Share
20:49
Then there is
machine learning
and the machine learning
has this incredible flexibility in the sense that you realize that when the machine learns by itself it finds all these database of this correlation etcetera, and makes a prediction based on that.Share
21:07
This is much more robust. This is much more flexible. Can have made some small mistakes, you can have a wrong data somewhere etcetera. It will still make the job because it averages of a large number of data. So there is this kind of law of large numbers that will help it to correct its own mistakes. So that is something that is a very striking of the new development.
Share
21:28
Now, how can one get an idea of that? I think one should be very lucid about what are these machines doing? People have tried to explain in terms which are not technical, how they work. I think it's very useful to open a little bit the black box.
Share
21:45
And I give a public lecture on
artificial intelligence
. I start with the two population of the colleagues, the ones who say it will change the world for the good and the one who say it will be a catastrophe and it will destroy the world.Share
21:57
Then I explain all the incredible new results that have been obtained, and the reading of the book of the life and the proteins etcetera etcetera. And then I have always a section opening the black box, I think that in some sense not going to do to the stage in which we are able to code an
AI
device. People who have a little bit of education and curiosity, they can have an idea of how it works.Share
22:24
It's relatively easy, you can imagine that you have a machine and this machine, it has thousands, actually millions of small buttons and these buttons, they will tell you instruction of the machine.
Share
22:36
The traditional way was, "I have to find the right direction for each of these buttons, there are many, many combinations and they have to think and find the program and that adapts all these things".
Share
22:46
And in
machine learning
is completely on the contrary, you start with your buttons in completely random position, it's a mess, does not do anything. You present image, it tells you a nonsense at the exit.Share
22:58
And then you start to say, "Well if I turn this knob here a little bit to the left, I present the image again, does it do a better job? It does. I leave it there, I try another button." And you do that and gradually after billions of tries it will do it. So I think that having this idea in mind is already something.
Share
Virginia Stagni
23:22
Thank you, thank you so much for sharing this. Now some of the challenges. So the students and early career professionals that took part in the past editions of FT Talent Bocconi challenge are coming to the show to ask some questions. So we have Jean and Helen for you. And Jean I know you are here with your questions so over to you.
Share
Jean
23:43
Hello, my name is Jean. I was a participant of the FT Talent Challenge in 2022. I grew up in
France
and Germany,
but I currently live in Singapore
where I'm working in Fintech on solving some of the industry's ISG data challenges. My question to Professor Mark Mazer is the following:Share
23:59
Growing capabilities of
AI
tools have the potential to redefine our occupations - think of Dall-E
2 and graphic designers for instance - what key strengths should humans develop from a professional perspective to find relevance in the new AI
driven environment?Share
24:15
Thank you. Professor Marc Mezard. I look forward to hearing your insights.
Share
Marc Mézard
24:20
Thank you Jean for this excellent question. As I was describing AI is presently reshaping the world and it creates a lot of interrogation to the young generation. What will this job still exist in the future or not? I think there is a kind of generic answer that I would like to give. I will not address specific cases let's say, but I think that one should be able to do.
Share
24:44
First of all is realized how important it is in present time to have a training that allows you to change orientation, to be able to modify your work program and schedule. And then the second point which is important also, is how to have this background, that is a background that allows you to understand how data can influence or not the job that you're doing.
Share
25:12
But if you can foresee that there is there a lot of data and there is the new oil - as it was mentioned by Virginia before - that in your field data is becoming very important, then that is a good signal and it's a signal that it's something that will maybe reshape your activity.
Share
25:33
I think that the young people like Jean, they are able to take it and become not just someone who suffers from it, but someone who is a motor of this evolution.
Share
Virginia Stagni
25:45
Thank you very much. And the last question is from Helen.
Share
Helen
25:49
Hello, my name is Helen Poon and I was a participant of the FT Talent challenge in 2021 and I live in
London
. I'm currently transitioning from investment banking to working in the group M&A and strategy team at the Financial Times
. My question to Professor Mark Mezard is:Share
26:09
With tools such as pseudo right available to students to write essays, and text image generators like
Midjourney
and Dall-E 2
available to artists, where should we draw the line between effectively utilizing AI
, and what is considered cheating? Should use of AI
be allowed in competitions of art and literature?Share
26:35
Thank you very much. I look forward to hearing what you think!
Share
Marc Mézard
26:42
What a nice question. Thank you, thank you Helen. First of all. Well, there are two things, there is a question of cheating, which is something that one should certainly avoid at the maximum. There is a good way to use these tools. The good way is to mention it, there is nothing wrong in sighting, writing even long pieces of papers that have been done before, provided you mention it.
Share
27:07
If you use them inside them without mentioning it, it becomes problematic. But if you look at a large part of literature, a large part of literature is built in reaction to previous literature, and often with the citations, which are not always explicit, I should say, but sometimes they are.
Share
27:26
And so it is a motor of the world. A motor of the world is to take pieces of what has been done before, incorporated in what you are doing, and then elaborating on it. Research works like that. Scientific research works like that also. We are building up on the papers of other people, that they have recently published, we have read them, it has inspired me etcetera. The point is a question of giving the correct quote again, there is absolutely nothing wrong in quoting someone.
Share
27:58
Now, if what you ask is is it fair to take an artificial program that will generate your essay and to give that to your teacher? This, I think it is absolutely clearly very unfair. It deprives you from the very nice experience of writing yourself an essay.
Share
28:20
And the point is not so much in general in the result, which is the essay itself, but the process of writing it which is the one that empowers you and that helps you to develop your skills.
Share
Virginia Stagni
28:32
Professor Mezard, thank you very much for today. Thank you for all your insights and thanks for sharing with us your journey as well as your expertise. I really hope you enjoyed our time together today. It was definitely a great learning experience for myself and I bet for all our listeners out there. So thank you very much. I cannot appreciate more your time.
Share
Marc Mézard
28:53
Thanks a lot
Virginia
. It was a great pleasure to discuss with you.Share
Virginia Stagni
28:59
If you're a listener of a Talent Show, I bet you're quite interested in the world of work, and in understanding trends that are shaking out workplaces worldwide today. I recommend you to check out Working It, the FT's workplace podcast newsletter. Join our friend and host, Isabel Berwick every Wednesday for understanding the big ideas shaping work today and the whole habits we need to leave behind, tune in, subscribe and follow.
Share
Add podcast
🇮🇹 Made with love & passion in Italy. 🌎 Enjoyed everywhere
Build n. 1.38.1
Marc Mézard
Virginia Stagni
Jean
Helen