0:00You said that it all started with a film 2001 a space odyssey. What was it about that film that stayed with you?
0:09Well, I saw this movie when it came out in uh in France in 1969. So, I was 9 years old and [snorts] it's the movie is about all the themes I was fascinated by at the time. uh space travel uh the the emergence of human intelligence and uh you know the next stage in human evolution uh and AI uh I mean the concept that you know computers could be intelligent and
0:38everything. So I was I was totally transfixed by that movie I think it changed my life.
0:45What did you imagine machines might one day do from that show?
0:52Uh well, I mean the concept that a machine could be intelligent, I think was in itself was fascinating. Uh the the movie was a little too optimistic about how long it would take [laughter] to get machines, you know, to the level of intelligence that they describe. Um, in fact, in the in the mo the movie takes place in 2001, but in the uh in the novel, the computer is is built in 1992. In 1982, we're
1:21nowhere near the the type of intelligence that we we see in Hal 9000 in 2001. So it took so so that that indicates something that is still with us which is that getting machines to get to the level of human intelligence is much more complex and complicated than we think it is.
1:40Uh and it's been a constant uh over over decades. The the scientist who advised uh Stanny Kubri uh and and Arthur T. Clark for a 2001 spacey was Marvin Minsky from MIT. He was a AI pioneer also a touring award uh recipient and uh he was too optimistic about how long it would take to get internal machines.
2:08All right. Thank you. So for years many experts said that neural networks were a dead end. What kept you going?
2:19Yeah, there were there were waves right.
2:21So in uh in the 50s the when the phrase AI was coined there were really two branches. What one branch uh was you know based on uh logic and reasoning and things like that and and led to uh what we now call good oldfashioned AI and to classical computer science. The other branch was based on an an an older field called
2:50cybernetics um which now would be called uh optimal control theory or something like this but um or pattern recognition and that branch is what led to neural networks. So, so there there was a wave of interest in neural networks and and machine learning in the late 50s early 60s and it died in the late60s because of a book that was written by Marvin
3:18Msky and Seymour Pepper both from MIT where they showed that uh the type of neural networks people were working on at the time were very limited in their capabilities and that kind of killed the field. Uh and so the there was essentially no work on neural networks during the 70s and during the 80s at least not in the west. There there was some work uh in Japan
3:45actually because Japan had a bit of a separate ecosystem for research [snorts] uh and and that got revived in the mid 80s uh with uh physicists essentially who got interested in neural networks because of some theoretical techniques that they thought they could use to explain how they worked. And then uh a new procedure to train neural nets called back propagation appeared in 199596. This is uh something I
4:15contributed a little bit to and it's still the technique we use today to train neural net. So there was a wave of interest between the mid 80s and the mid 90s in neural nets. Uh and then it died again in the mid 90s. Uh and the reason are not entirely clear. It's it's a set of methods that are really complicated.
4:35They require a lot of black art to make them work. And the way this black art is disseminated today is through open source code. But back in those days there was no internet.
4:49And people did not have the habit of uh distributing their code in open source and so uh it it basically made it difficult to make progress. So people kind of in the mid 90s working on on machine learning were working on simpler methods than neural nets. And then it took until the early 2010s for neural nets what we now call deep learning to take off again. They
5:17revolutionized uh speech recognition around 2010, image recognition around 2013 and then natural language understanding around 2015 16. And now with you know LLM uh they just completely took off.
5:36Thank you for your lesson on the history lifespan of uh neuronet networks. Now at Bell Labs your work help machines read handwritten chats. What was it like to see your research out in the real world?
5:52Yeah, so that was very satisfying uh to and and it's something I've tried to do all my career. it's uh you know work on a very fundamental problem which for me was how to get machines you know to be as intelligent as humans and animals. Uh and then very early on deciding that that had to involve learning because in nature every intelligent animal learns. Anything that
6:21has a brain learns uh and even things that don't have brains adapt. So uh so you know it's it's really very basic. Uh so that's what got me interested in machine learning very early on and then of course getting inspiration from uh from from from from biology but uh but also the fact that perhaps I didn't think people were smart enough certainly I wasn't smart enough to just conceive
6:50and design an intelligent machine from scratch like it had to build itself right so that was kind of a bit of my motivation is tenacity for you to to to to do all the research and making it come to life bigger than life and and so you start with this you know scientific uh graduous question and then you you try to take the first step and almost every time when you do fundamental research the the first step you you make in that direction turns out
7:18to to have usefulness. So my next step was you know after this idea that you could train neural networks with multiple layers which we now call deep learning using the back propagation algorithm which I I mentioned earlier. Uh I I got inspiration from the visual cortex the architecture of visual cortex to design a type of neural net called convolutional neural networks.
7:41I call them connet people call them CNN. I don't like that term because it sounds like a TV station, but uh but that is the type of architecture that allowed uh to to to do things like recognize handwritten uh you know handwritten characters and read checks, read zip codes, read envelopes, read documents.
8:04And so it worked. And then uh we worked with engineering group that turned it into a real product that uh you know banks were were using in the late 90s to read checks.
8:18This is very amazing interesting. So every scientist has people who shaped them. Who were yours and what did they teach you?
8:28Uh well my scientific hero is Albert Einstein. [laughter] uh but uh so I'm in that sense I would be a failed physicist but it's because I I have the sort of engineering uh you know the the the idea of like you know building things at heart was an engineer too and he taught me a lot about you know how you build things how you design new things you know that nobody
8:57thought about before and so this this creative side of thing I think was fascinating for me So that's why I went into engineering and not into theoretical physics. Uh but uh but my intellectual hero was uh Einstein. Another one is uh Richard Fineman also a physicist.
9:18We have spoken before about creating machines that can reach human level intelligence. What do you think the role of humans could be in such a war?
9:29So a lot of people have a fear of what's going to happen when you know machines start to be as smart as humans. Uh my opinion is that it will elevate humans. Humans will be the the boss if you want of teams of AI staff. You could think of it this way. All of us would be a boss. Okay? And the AI will amplify your intelligence, our creativity, our efficiency.
10:03Uh and uh you know, I think we are entering a we'll be entering a new a new era. Uh, and I don't think people should feel threatened by, you know, working with entities that are smarter than than them any more than say a political leader should be threatened by hiring staff that are smarter than him or her [laughter] or or or you know a lab director like me should
10:31be threatened should feel threatened by you know hiring students and scientists who are smarter than himself. That's actually the best thing that can happen to you. So, so that that's the way we we should think about like we will serve as uh a direction which means that [snorts] um people will will need to have like more of a moral fiber and like know what's useful, what questions to ask.
10:59The difficult thing in science and engineering very often is not to answer a question, is to figure out what is a good question to ask. And uh I think humans are still going to be doing this to a large extent.
11:11I like that answer. Yeah. Um setting boundaries and becoming more embracing to technology. Yeah. I like this question. There's so much miscom misinformation on AI out there.
11:27If you could just correct one misconception, what would that be?
11:32Oh, there's a lot of mis misconceptions.
11:34Okay. The first mis misconception is that we're going to have super intelligent system by next year or two years from now. That's not going to happen. Uh it's going to take much longer. What we have already are AI systems that are extremely useful. They can answer a lot of questions. They're not particularly smart. They just have a lot of accumulated declarative knowledge so they can regurgitate it at the right time. Uh we have systems that can do specific tasks really well like proving
12:02math theorems uh generating code uh predicting the structure of proteins you know playing chess and go you know things like that right and and the list is going to increase uh with time um but ask yourself why we still don't have domestic robots or um autonom completely autonomous self- loving cars at at level five. which don't have that. And it's because uh
12:31our systems, our AI systems are really good at manipulating symbols like mathematics and computer programs. They're terrible at understanding the physical world. They they're not nearly as good at understanding the physical world as a cat. Right? So, we still have a long way to go um for what some people have called physical AI. AI that understand the physical world. That's what I'm working on.
12:59Also in my new company, Emmy Labs and uh it's still a big challenge.
13:04Very interesting. Thank you for your enlightenment.
13:08Um outside the lab, you build model planes and uh play jazz. What do they give you that research doesn't?
13:19Well, I like building things, you know.
13:21I like coming up with new ideas and building things. nobody has thought about before. And my my dad was a mechanical engineer uh working in the aeronautical industry. So he got me into building airplanes, model airplanes. Uh it become kind of a a family hobby. My brother is into it too. Um [snorts] and and so I find I find that act of like building something with your hands and getting it to fly really fun.
13:50[snorts] Um, but I'm also into music and, you know, a few other things. So, you know, other creative activities.
14:00I think it kind of relates to to to your work. You're a builder. You build things.
14:05Yeah. One of the things that uh I I I taught myself electronics when I was a teenager. And one of the reasons is u I wanted to build electronic music instruments like synthesizers and stuff like that. So, [laughter] okay. You kind of deviate from where where where you first wanted to do.
14:22I was also interested in, you know, intelligence, but I was naive about it at the time.
14:28Naive. Okay. You often tell young people to make yourself useful. What does that mean for a student today?
14:37Oh, a student should should not try to make themselves useful. they they they should try to learn uh as much fundamental topics as possible particularly topics that have a long shelf life. So things like you know uh basic courses in in physics or engineering uh science or mathematics um in in the sciences uh because those
15:04won't won't go out of fashion. Um, I advise students to not take too many like technology courses that are because the type of technology you you learn as a as a short shelf life is very likely to be out of date five years from now. And because technology progress technological progress accelerates, uh, young people coming out of college are going to have to learn new things constantly.
15:37and almost certainly their profession is going to change you know during their career. So they should prepare themselves to learn they should learn to learn and the the best way to do this is to learn very very basic fundamental things that have a a long shelf life. So the joke I I say which I I said in my talk today is if you had the choice between taking a course on you know iOS programming or quantum mechanics take quantum
16:07mechanics because you will learn basic methods that will help you understand new things that will pop up five years from now that you may not be equipped to understand otherwise.
16:17I love your advice. I if students are hearing this they should remember this question. Now I know why when I was schooling my my my students as my my teacher was asking study your math, practice your math because it's a core is actually the core uh subject I think to a lot of things.
16:37Uh you've said the biggest questions in science are about the universe, life and the brain. If you were 16 again, which would you chase?
16:48Yeah. What what's the universe made of?
16:51What's life all about? and how does the brain work? Those are the three basic big scientific questions, right? I don't know. I'm fascinated by all three of them. Uh I think life and biology is really messy. So uh you know the kind of science you have to do there is uh not the kind of science that I think I'm I'm good at because a lot of it is uh experimental.
17:15Uh and then you know how does the brain work? like do you need to study the brain or do you need to uh like come up with theories and algorithms? So as a as an engineer I think the best way to validate some theory about how the brain works is to actually implement it on a computer. Uh but you know the same way that uh you know the early pioneers of aviation were
17:45inspired by birds but they did not reproduce a bird. Some of them did but that was a mistake but uh but they got inspiration from bird and then derived from underlying principles about how birds fly and then built flying machines relying on the same principles. So I think the the science of intelligence if you want is should be the same. You study human intelligence try to derive
18:14some underlying principles like learning uh etc and try to reproduce this with machines and the best validation for whether your theory is is right or not is whether those machines exhibit intelligent behavior. So that's why in my career I've been always inspired by neuroscience and cognitive science.
18:39Very interesting of the relationship that you make between uh them. Final questions. If you could give one piece of advice to a young person who dreams of becoming a scientist, what would it be?
18:54Well, you have to imagine what the hot topics in science are going to be. uh let's say five six seven years from now uh and uh and how science will be conducted. So clearly AI is going to be used massively in the process of research not just to generate code but also to uh produce experimental um experiments and uh perhaps automate
19:21some of the experimental processes in chemistry lab for example. Um, I think there's a lot of discoveries that are where AI is so promising that you would need scientists that know the the domain science that they are interested in, but also know sufficient AI that they can build themselves tools that no one else has based on AI. Right? Because a lot of
19:50discoveries in science have been brought about by the construction of new tools.
19:55the telescope, you know, allowed Galileo to, you know, discover like major uh things about, you know, how how the how the world works. Uh, you know, same for the microscope. It opened the door to a lot of discoveries. So, so it's very often an experimental technique or a new instrument that brings a kind of a new era in science. AI is going to bring a new era in things like material
20:24science, biology, medicine, and you know, all kinds of uh other areas uh environmental protection and things like that. So, I think it's you know, incredibly exciting. We're going to see an acceleration of progress. And what that means is that again uh young people should learn very basic fundamental topics because their knowledge is going to have to evolve with technology which is accelerating.
20:51So it all goes down to basic back to big.
20:56Thank you for your time uh today. So with that we have come to the end of the interview. It was a pleasure. Thank you.