AI pioneer Yann LeCun: India must embrace open source, invest in research to become an AI hub like France

It has been a hectic week for French-American scientist Dr Yann LeCun, who with Geoffrey Hinton and Yoshua Bengio, is often referred to as one of the Godfathers of AI, for his pioneering work in deep learning.

LeCun, who is VP and Chief AI scientist at Meta, often gets mobbed for selfies these days, amid the ongoing generative AI boom.

Unlike Hinton and Bengio, LeCun is optimistic about the benefits of artificial intelligence and calls out any harm arising from it as ‘science fiction’ at this point. He is also a firm votary of open-source research and believes India should build on top of an AI model rather than build a large language model (LLM) from scratch.

He spoke to Moneycontrol on the sidelines of Meta's 'Build with AI' summit in Bengaluru. Edited excerpts

This is your third visit to India and your second to Bengaluru. What has changed and what hasn’t?

So many things. The first time I came to Bengaluru was in 2011, and this was before the deep learning revolution. I went to a scientific workshop at the Tata Institute nearby, and there was a big debate in the scientific community about where AI was going to go, and it was before people were really excited about deep learning.

Then I came back in early 2020, just before the pandemic, and there, AI had taken off, but mostly the applications were in vision, computer vision, speech recognition, speech synthesis, a little bit also in natural language processing, but the idea of LLMs and chatbots hadn't really emerged yet.

There was a lot of excitement from students and the technical academic community and, to some extent, the tech community about AI. But today, it's completely exploded.

Now I walk on the street, or just stand in the lobby of a hotel for a few minutes, and people come to me and say, oh, Professor LeCun, can I take a picture with you? I have a startup working on AI. I'm using Llama, the open-source model that Meta distributes. So, it's been completely exploding in India and in other countries as well.

Earlier you were on stage with Nandan Nilekani, where he said India should not be wasting money on building LLMs but focus on use cases instead. What do you think India’s approach should be, because there is a sense that India is simply building on top of foundational models being built abroad. That we need our own model for sovereignty.

Probably the right approach is to use an open-source foundation model, and Llama 3.2 now is the best one to use and find your need for a vertical application, find your need to speak all the 700 languages in India and have voice-activated systems so that everyone can have access to it. There are a huge number of vertical applications in the consumer and business domains that can be deployed.

Medium term, what will happen is that the training of foundational models is going to be distributed among several regions, the reason being that the cost of training models is ballooning. That's the first question.

The second one is that we need to have foundational models that speak all the languages in the world.

A single entity like Meta basically does not have access to all the data in the world, in part because regions want some sovereignty over their own data and control. Then the third one is also regulation. Some countries don't want the data to go out for various reasons.

So I think where we are going is a world in which AI is going to be essentially a shared platform, which is going to be trained in a distributed fashion and in such a way that the best open-source AI models will essentially take into account the entire knowledge of humanity, if you want. And then on top of this, those systems will be fine-tuned for multiple applications. I think the base models will be much more general than they currently are, and much less biased towards English.

Is AI becoming a race to the bottom? The price per token of generated LLM output has dropped 100X in the last year. Will selling intelligence become like selling rice?

Yes and no. It's going to be a commodity the same way as the evolution of the software stack of the internet or the mobile communication system.

25 years ago, the software stack was mostly proprietary. You had suppliers like Microsoft and Sun Microsystems and Oracle providing the basic components for software, operating systems. It all disappeared. Now it all runs on open-source platforms. Linux, MySQL, Apache for web servers.

The entire internet runs on open-source. The mobile communication system runs on an open-source stack. Not a widely known fact.

Your car runs Linux. Your phone probably runs Linux. So I think it's going to be a similar phenomenon because AI is going to become a common platform, it will need to be open-source. This is really the direction of history. I see that as inevitable. There will be lots and lots of contributions to it.

What you need in addition to this is a race to the bottom of the cost of running it. Because if you want very wide access to AI assistants, there is a future in which we're all going to be wearing smart glasses and be talking to our AI assistants, including people in rural India. They're ploughing their field and they're wondering, should I plant now or wait? They will be asking their assistants.

You also mentioned that if you want human-level intelligence, don't focus on building LLMs and abandon generative models. Why is that?

LLMs are wonderful, super useful, but they, in themselves, are not going to take us towards AI systems with human-level intelligence. They're very good at manipulating language. You can have an LLM pass the bar exam or write an essay.

But where is my domestic robot that can learn to clear out the dinner table and clean up the dishwasher without requiring hundreds of engineers engineering the system into it? Where is my self-driving car? Where is a system that can learn to drive a car in 20 hours of practice, like a 17-year-old or an 18-year-old?

The system can just, like a 10-year-old, come up to the table and figure out what to do. So that kind of intelligence, which requires an understanding of the physical world, does not exist yet. This is what we're building at FAIR (Fundamental AI Research Lab). We're working on the next generation AI systems, beyond LLMs. Systems that understand the world, have persistent memory, can plan and reason. Current LLMs can’t do this. This is a big challenge for the next five years to a decade.

You famously said that a domestic cat is smarter than AI. But do we really need AI to be as smart as a cat or a child? Because even with the developments we have now in AI, it's enough to create content, replace people doing customer support, analysing health records. It's already displacing jobs to some extent, at least at the lower levels. So, what's the argument for a human-level AI? And are we prepared for that kind of world?

It's displacing jobs, but it's mostly transforming jobs? It's making everyone who has a job more productive in that job. And then changing the focus of the activity that those people are doing.

But it's not really making jobs disappear. It makes new jobs appear, and it's very difficult to predict what the hot jobs 10 years from now will be, because of AI in particular. So we need human-level AI systems because we need assistants that are with us at all times. They can help us in our daily lives. They can solve problems for us.

For the same reason, if we are a leader in business, academia, or government, we need a staff of people helping us. We can't solve every problem. We have sort of limited brain cases. It's very empowering to be working with a staff of people who are smarter than you. So eventually, we're going to have AI assistants, virtual people, who are smarter than us, and it's going to be empowering for us. We're going to be calling the shot. We're going to be their boss and tell them what problems to solve for us, and they will be solving those problems for us.

It might have an equalising effect. We're seeing that people who are really good at their job, they're getting just a little better with AI, and people who are not so good at their job get much better with AI. It might have a funny effect of equalisation effects.

That's why it's so important also for a wide diversity of people to have access to AI. It's also very important for the AI systems to be diverse so that people can choose what type of AI system they want. You don't want just to have a choice between 2-3 proprietary AI systems that come from the West Coast of the U.S.

What role can India play in building AI models with more diversity and sensory inputs? We have a large dataset of languages, sub cultures and a huge digital public infra that is powering everything from identity to payments to delivery of the govt’s welfare schemes. 

In terms of the type of digital infrastructure that India has built, with very low-cost 5G access, it's pretty impressive.

Authentication technology, the substrate is already there, like the people who have the expertise, they need to provide this at a large scale. So there is a role that developers in India are already playing, which is essentially fine-tuning and adapting open-source platforms for vertical applications.

I think, the way India should migrate its influence is taking a bit of a bigger participation in the research community. Not just engineering and product development, but also research. I think it's important, because we've seen that effect in other countries, in France in particular. We created a research lab in France 10 years ago, in Paris, and this had an enormous effect on the local ecosystem.

It's become the second-biggest AI hub.

Exactly. And it's because we created that lab. The psychological effect was enormous. They said AI is the cool thing now. I have potentially a career staying in France because of the existence of this lab. Our lab actually prompted Google to create a lab and then French companies to embark on the same thing. So it created an ecosystem and career prospects.

It motivated students to, instead of going into finance, to do a PhD in AI. Those are students who spend time at FAIR, basically co-advised by people at FAIR and universities. Some of those people we hired at FAIR, but most of them went into the ecosystem and did startups. It totally lifted the ecosystem.

I think India can do the same thing, probably centered on Southern India - Bengaluru or Chennai (IIT Madras), and perhaps others.

You are wearing Meta glasses. I was curious, what do you use it for on a daily basis? And how do you see AI kind of evolving form factors over the longer term?

I can take pictures with them or I can ask you to take a picture. Or you can take videos. But you can also be looking at a menu written in Kannada or something, and it would translate it for me.

I walk to work. Most of the time, if I walk on the street, I play music through that, and I can play messages.

You're very different in your view of AI from your fellow Turing Award winners, (Geoffrey) Hinton and (Yoshua) Bengio, in terms of the adverse effects of AI. I mean, they're more on the AI doomerism side of the debate, but you are an AI optimist. Why is that? And don't you think there should be some level of regulation, or a balance?

There is certainly nothing wrong with regulating the deployment of products, whether they use AI or not. If you're going to put a driving assistance system in a car, you probably want some government agency to make sure the system is safe, regardless of what other technology is used for it. Or if you have a medical image analysis system, which may be based on an AI technique, you want that to have some sort of clinical trial and be vetted by the government.

So there needs to be regulation for products, and most of them are already in place for life-critical applications.

The people who are in favour of research and development regulation are people who think that AI is intrinsically dangerous, even if you do not build products out of it. It is built around the idea that somehow if an AI system is built that is sufficiently powerful, it could fall in the wrong hands by accident. Or just the mere fact of turning it on could be dangerous for humanity. I do not believe in this at all. I think that's science fiction. I think it's by people who really don't know what they're talking about. Or people who have inflated ideas of what AI systems can do today. So I don't believe in this.

Now, my colleagues and friends,Bengio and Hinton, it's different. They don't necessarily have as much confidence as I have in the ability of democratic institutions to take the best advantage of technological evolutions. They're a little bit more negative or skeptical. I mean, they see what's happening with climate change and they say, if everything is motivated by profit, it's going to be bad. Things are going to happen just because the motivations are not the right way. I'm much more optimistic than they are. I think it should be in the hands of democracies to figure out what to do with it, but not regulate R&D.

Certainly, not putting any obstacles in the dissemination of open source AI platforms because that would be very dangerous. It would lead to regulatory capture and the fact that the number of providers of AI systems would be a small number of large companies. You don't want that to happen. You want AI to be a very distributed, widely available open source infrastructure.

How far are we from AGI or AMI (advanced machine intelligence) as you call it?

AMI. I think we're going to start to see some effect of the ideas that I talked about earlier. Maybe within three years. If things work out, we're going to be on a good path towards machines that can learn from the real world a bit like animals within five-six years, maybe seven. And then perhaps human-level AI systems within the decade. But there is a long tail on the distribution, which means that it could take much longer because we could be hitting obstacles that we're not foreseeing.

It's almost always the case in AI that a new technique appears, a new paradigm or a new set of methods. And the community says, that's it, now we're going to get to human level AI. We just need to scale it up. That's what people are thinking about LLMs and it's wrong.

That's what people were thinking about the previous generation of AI techniques and that was wrong. So maybe I'm wrong too. There is a probability for things to be much harder than we think.

As things stand today, what would your advice be to a 10-year-old, a 20-year-old, and a 30-to 40-year-old who's seeing AI everywhere, getting overwhelmed in the process because the pace of disruption is hard to keep up with?

A 10-year-old is not going to have any problem because that 10-year-old is going to grow up with AI. And for that 10-year-old, AI would be part of the scenery. It would be completely natural.

For a 20-year-old, the question is what specialty should I learn? Is it worthless to learn computer science because AI systems are going to programme better than they can? No, that's not true.

You should study very basic things that have a long shelf life - mathematics, physics, basic computer science, applied mathematics. Those are things that would be necessary to understand and develop the next generation of AI systems.

If you had a choice between taking a course on mobile app development and learning quantum mechanics, learn quantum mechanics. That's my recommendation. Learn things that have a long shelf life.

For people who are 30-40 years old, the world is going to change a lot. Don't put all your chips on what you think is going to be the big thing because it's going to change within three to five years. Technology is going to change completely. The capabilities are going to be much bigger than they are currently. And so don't make choices that sort of make you prisoner of a hypothesis about where technology is going.