Speaker A
You know, what is the right thing to work on. That's the, you know, what is going to work, what is going to make you, you know, work towards, like, significant progress, which may be very long term.
Yann LeCun discusses AI, career transitions, and lifelong passion for intelligence research in a candid conversation with Nitin Dua.
Key Takeaways
- Working on the right problems is crucial for impactful AI progress.
- Humility and awareness of one’s limits foster better collaboration and innovation.
- Conceptual and perceptual advances are as important as mathematical or coding skills in AI.
- Career transitions between academia and industry can enrich research and application.
- Lifelong passion and mission-driven work sustain long-term contributions to AI.
Summary
- Yann LeCun reflects on his career journey from industry research labs to academia and Meta AI leadership.
- He emphasizes the importance of working on the right problems for long-term significant progress in AI.
- LeCun shares humility about intelligence, acknowledging many peers smarter than himself.
- He highlights his focus on conceptual advances in machine learning, particularly perceptual advances.
- LeCun describes early work on machine learning through optical character recognition and convolutional nets inspired by biology.
- He discusses the evolution of neural networks research and its initial rejection by the community.
- LeCun talks about balancing roles in academia and industry as complementary and fruitful.
- He recently transitioned from Meta to founding AMI Labs, focusing on scientific leadership rather than operational management.
- LeCun’s mission is threefold: discovering intelligence mysteries, building intelligent systems, and applying AI broadly.
- The conversation touches on his views on religion, atheism, and causal inference.
Full Transcript — Download SRT & Markdown
Speaker A
And despite the fact that my wife wants me to retire, I just, I just feel I have to kind of push this, push this forward.
Speaker A
So this is quite fulfilling to you. You know, you spoke something in between where you said you initially felt you were not smart and you felt others also are not smart in that sense.
Speaker A
Well, there's a lot of people who were smarter than me and even met, you know, friends who are better than me at both.
Speaker A
And that's really humbling. If you could have someone like that, like you want to work with those people.
Speaker A
I think we are the easiest ones to fool ourselves. And if we can start seeing that crap, it's a great state to come in.
Speaker A
If you can arrive at that level of awareness, in my word, operate with that consciousness, I think it's a wonderful place to be.
Speaker A
It's time to kind of go into high gear and try to make this real and applicable to a wide range of applications in the world, in biomedicine, in industry.
Speaker A
What's kept this underneath fire alive for you to constantly give to this space and constantly have this fire to learn?
Speaker A
So I think my mission in life, yeah, is threefold. Discovering the mysteries of intelligence, which is a scientific question.
Speaker A
And the best way to do this is to build intelligent systems. Right. So that's my engineer side.
Speaker A
So here is the offensive statement. If humans were so good at causal inference, religion would not exist, because religion basically, in part, is.
Speaker A
In this video, I talked to Yann LeCun, who's a former chief AI head at Meta.
Speaker A
His career transitions at 65 years. He's transitioned to building AMI Labs, raised a billion dollars for it.
Speaker A
I get into what's kept his fire alive for all these 40, 50 years of doing the work in deep learning, including winning a Turing Award.
Speaker A
We get into his perspectives and beliefs. He also talks about religion as an atheist, and I have a different take on it, and I respectfully disagree on a few things with him.
Speaker A
I hope you find value in this conversation. On his work on AI career transitions, fulfillment and.
Speaker A
And different perspectives. Enjoy this video. This is Nitin Dua serving you with love. You know, you spoke something in between where you said you initially felt you're not smart and you felt others also are not smart in that sense.
Speaker A
So. Well, there's a lot of people way smarter than me. If I can, if I can find them, I hire them either as collaborators or students.
Speaker A
No, absolutely. But just want to go deeper on that. There is also this thought that, I mean, I come across executives who are terrifically smart but also come to a point where they feel they're not good enough in what they're doing.
Speaker A
Has that happened to you? Maybe not now, but at some point. Every time. Every time, yeah.
Speaker A
I mean, I've been working with, you know, friends and collaborators that are like much better, much better than me and pretty much all the domains that I think I'm good at.
Speaker A
So I mean, I've met people who are like way better than I am at theory and the mathematical aspects of machine learning or other things.
Speaker A
People that are much better than me at coding. Certainly now I'm older, so I'm less good and even met friends who are better than me at both.
Speaker A
And that's really humbling. If you could have someone like that, you want to work with those people.
Speaker A
No, I think what I've been trying to sort of contribute to is sort of perceptual advances, right?
Speaker A
So it's not necessarily kind of the mathematics or the implementation, but it's the thing in between which is like conceptual advance.
Speaker A
What is the right thing to work on, what is going to work, what is going to make you work towards significant progress which may be very long term.
Speaker A
For example, sometimes my earlier work is described. Yann's work is in optical character recognition.
Speaker A
I never had any interest in character recognition. What I was after was getting machines to learn.
Speaker A
And the first example in 1980s of what you could do with machine learning was twofold.
Speaker A
Either you were going to do something like character recognition. You have to realize 1983 or so, it's not like you could just plug a camera on a computer and grab images.
Speaker A
So there's no. It was very complicated to get images into. I remember working with Dr.
Speaker A
Bernie on the character recognition part. Back in the day there was this project that he's working on on pattern recognition.
Speaker A
And that's what we started, right? How can we make the machines smarter on learning that as you.
Speaker A
But there was no Internet, there was no cameras you could use, right? So there were like, you know, there were like industrial scanners, there were line scanners, right?
Speaker A
So there was some data sets that you know, you could get on tapes with characters, right, from the post office or from, you know, bank checks or something like that.
Speaker A
And the other domain where you had data was speech recognition. So I focused on characters because that was the only thing that you can get data for which was sort of perception.
Speaker A
Vision related speech I thought required more specific approaches. So I didn't work on it.
Speaker A
Some of my friends did. That's why I used characters. It's not because I was interested in the problem.
Speaker A
I was interested in making progress with machine learning, machine perception, what we now call deep learning, which means training the system to not just classify but to learn to represent and extract features in a hierarchical way.
Speaker A
And that's what led to convolutional nets. I was inspired by biology in doing this and got some really good results in the late 80s, early 90s, actually deployed some commercial practical applications for reading checks in the mid 90s.
Speaker A
But by then the research community had no interest in neural nets anymore. And so then it took another 15, 20 years to kind of convince the community that those methods actually were working.
Speaker A
Okay, so I see still one thing here, that your constant interest, you've transitioned different roles, right?
Speaker A
Maybe you played the academia, you've also been in industry at Meta. Well, I spent the first 12 years of my career as in industry research labs at Bell Labs, AT&T Labs, then the NEC Research Institute for 18 months.
Speaker A
Then I became a professor in 2003. I was 43. Okay. And for about 10 years was just a professor, you know, help co-founding a few companies, but that was not my main livelihood.
Speaker A
And then in 2013 joined what was then Facebook and started the industry research lab.
Speaker A
And I thought I knew what I was doing because I had worked at Bell Labs and sort of got inspiration from that, but maintained my position also at NYU.
Speaker A
So I had a foot in industry, a foot in academia. And I think that those two worlds are complementary.
Speaker A
And so I find this dual affiliation really very fruitful. Yeah, no absolutely. But I think most recently you've transitioned from there from Meta and now I think you have a lab.
Speaker A
So now I'm starting a company called AMI Labs, Advanced Machine Intelligence Labs and I'm the executive chairman.
Speaker A
I'm not the CEO because I'm not an operational person and not a particularly good manager.
Speaker A
So I'm more into, you know, kind of, you know, science and technology, strategy and direction, scientific leadership and things like that.
Speaker A
So, so that, that's the role that I had at Meta and which I'm going to continue having.
Speaker A
I also keep my position at NYU. Yeah. To have the confirmatory, what's kept all these years, I mean, you've given at least four decades to this kind of work.
Speaker A
I mean, it's been progressing. Some of it, as you said, have been repurposed, relabeled, but constantly upgraded to understand and observe what's happening around us in reality.
Speaker A
What's kept this underneath fire alive for you to constantly give to this space and constantly have this fire to learn?
Speaker A
So I think my mission in life is threefold. Discovering the mysteries of intelligence, which is a s
Speaker A
And the best way to do this is to build intelligent systems. Right? So that's my engineer side, to verify that the hypotheses you made actually are actually true.
Speaker A
And then the third one is, having an impact, positive impact on society with that kind of technology and so pushing it into the real world, which as a scientist you can do either purely intellectually through papers and demonstration and software and
Speaker A
things like this, or you can push to the next step and do technology transfer or, you know, work in the, industry to kind of try to push this in products or create a startup to really kind of push this.
Speaker A
Our team requested to share this with you. Our data suggests that many of you who have been watching our videos have still not subscribed.
Speaker A
Some of you have reached out to me to know when the next video comes out.
Speaker A
May I request you to please subscribe and hit the bell icon so you can get notified when the next video goes out.
Speaker A
By subscribing, you also become part of our mission. Empowering professionals, helping progress in the careers, integrating health and mind and fulfillment.
Speaker A
Please share this with someone who may find value in this. You know, in the last five years, we've made some advances and discoveries on how to really push AI technology forward.
Speaker A
You know, I could have decided five years ago that I was starting to kind of, you know, lose it and maybe I was too old or whatever and, lost creativity or, or maybe I didn't, you know, I ran out of
Speaker A
ideas. But it turns out like, you know, I have this new avenue, which I think is fascinating.
Speaker A
And despite the fact that my wife wants me to retire, you know, I just feel I have to kind of push this forward.
Speaker A
So. So, this is quite fulfilling to you, to constantly push the boundaries to understand the mysteries of the universe.
Speaker A
That's right. And we've made progress in the technology of, you know, building AI systems that we think can understand the real world and to the point where it's time to kind of go into high gear and try to make this real and
Speaker A
applicable to a wide range of applications in the world, in biomedicine, in industry and all kinds of corners of life and ultimately build systems that amplify human intelligence, which really I think is a intrinsically good thing to do.
Speaker A
Yeah absolutely. I think there is a lot of merit to doing that. But in all of this, I mean as I was sharing with you Yann earlier, that you know, I go in gather perspectives like you have shared
Speaker A
here, but also from other facets of life, like you know, neuroscience, you know, even spiritual leaders and many other formats.
Speaker A
Yeah, I'm a completely non spiritual person. I'm like you know, extreme rationalist. I'm you know, an atheist and everything.
Speaker A
And I don't think there is anything particularly mysterious or magical about the human mind or consciousness or anything like this.
Speaker A
So like. Yeah. Okay. And do you think to that end the science and the rational side of it understands what's happening in reality?
Speaker A
I mean, giving you so much time to it? Yeah, yeah, that's it's part of the scientific path that I'm following.
Speaker A
Yeah. And what makes you not believe in any of the other stuff? just curious to hear that. which stuff in particular like religion, something that science does not yet understand and cannot box down to.
Speaker A
Eventually we'll understand it. Yeah, I mean there's been a history of humanity, Right, that you don't understand some mysterious mechanism in nature or human nature, and you attribute it to, you know, mystical forces or underlying divinities of things.
Speaker A
And the history of intellectual progress has been to push this away. The more we understand about the world, the more we realize, okay, there is no God that controls thunder or something, Jupiter or whatever.
Speaker A
Well the name could change. And again, people can have different perspectives on it. That's just Yann's perspective.
Speaker A
Okay, I'm going to be extremely obnoxious and offensive. There's a lot of people in the context of AI and other sciences who are talking about causal inference.
Speaker A
So causal inference is the problem of you're observing a phenomenon and you're trying to figure out what is the cause of this phenomenon and can I influence it by changing the cause.
Speaker A
Right. And humans are pretty good at this, but they're very far from perfect, particularly little children.
Speaker A
So little children, you ask a lot of 5 year olds what is the cause of wind?
Speaker A
And half of them will tell you wind is caused by the motion of leaves on the trees.
Speaker A
They get the causality backwards. Right. Okay. It's the other way around. Right. It's the wind that causes the leaves on the trees.
Speaker A
To move as we know. But they don't have a good enough mental model of reality to really kind of realize what they're saying is not possible.
Speaker A
So they come up with an explanation. And in the course of human history we've come up with this type of backward causality or inventing causes for things that really don't have direct causes.
Speaker A
Right. So what causes thunders? Well, there's a God that decides to kind of put a lightning bolt on your house and destroy it because you are a bad person or something.
Speaker A
So here is the offensive statement. If humans were so good at causal inference, religion would not exist.
Speaker A
Because religion basically in part is attribution of causes that are wrong. And science basically keeps pushing the frontier and pushing divinity in the gaps that we don't understand.
Speaker A
But eventually we'll figure it out. Yeah, I mean, my view on this, I mean, it's a whole day topic that we can spend time on.
Speaker A
And what I will say here is I think there's a lot of layers to it.
Speaker A
Also perspective matters. Also what we see and what we don't see. As you said, observation is super important.
Speaker A
So I think it's also a lot of experiential that you can't that science today cannot explain fully.
Speaker A
And I think even religion. So there are a bunch of narratives. I mean, I definitely feel like religion first is quite different than spirit.
Speaker A
And this is me speaking from a prior science lens. But I also see its limitation, at least in my experience and my view.
Speaker A
But I totally respect different perspectives on it and implicitly. Yeah absolutely. And I think that's why we have these discussions because ultimately the mystery, ultimately, you know, if we can get to that is just so magical.
Speaker A
If we do well, we need to, uncover mysteries. That's the point of the scientific method is, you know, uncover mysteries without fooling yourself.
Speaker A
I mean, the whole scientific method is to figure out how to use empirical evidence without fooling yourself from your own biases and absolute preconceived ideas.
Speaker A
That's a beautiful. I think we are the easiest one to fool ourselves. And if we can start seeing that trap, it's a great state to come in.
Speaker A
And I think however you arrive at it, whether through science or through another endeavor, if we can arrive at that level of awareness and not fool ourselves any longer, and in my word, operate with that consciousness, I think it's a
Speaker A
wonderful place to be. Thank you Yann, for sharing your perspective on it. And this is Nitin Dua serving you with love.
Topics:Yann LeCunAI researchmachine learningdeep learningMeta AIcareer transitionconvolutional neural networksintelligenceAMI LabsNitin Dua
