And now to introduce our final genius honorary of the evening, let me welcome to the stage LSE trustee Lindsay Hans. She is president and co-head of Merrill Wealth Management.
Good evening. Thank you, Paul. Yann LeCun is the executive chairman of AMI Labs and a professor at NYU. He was the chief AI scientist at Meta, as well as a founding director of Meta FAIR, and the NYU Center for Data Science. Yann is regarded as one of the three godfathers of AI. He did pioneering work in machine learning, computer vision, robotics, and computational neuroscience. Yann believes that the current large language models, which are favored by OpenAI, Google, Microsoft, and Anthropic, will not get us to general machine intelligence. He is pioneering an alternative to all of that: world models. It's an approach to AI, and he also co-founded the startup AMI Labs, which is backed by more than 1 billion of funding. He is a recipient of the Turing Prize, which is the Nobel Prize of computer science. He says that his interest in machine intelligence dates back to his childhood fascination with HAL, the murderous mainframe in 2001: A Space Odyssey. To protect people from rogue AIs in the future, Yann says that we must build into machines the notions of empathy and submission to humans. Paul, can we queue the video?
Ask yourself, why do we have those AI systems that, you know, we can pass the bar exam, push your arms, generate code. But then, where is my domestic robot? Where is even a cat-like robot that is as smart as a cat? Like, we don't have that. We're not even close. Yann LeCun is one of the most celebrated and contrarian minds in artificial intelligence. As Meta's chief AI scientist and a Turing Award winner, he helped build the foundations of modern AI. And he thinks most people have it fundamentally wrong. We have systems that can deal with sequences of discrete symbols like language, but that turns out to be simple. The real world is messy, and that's much more complicated. LeCun grew up in the suburbs of Paris in the 1960s. The son of an aerospace engineer who built things for the joy of it.
My dad was a mechanical engineer who worked for the aerospace industry. He taught himself electronics. He built airplanes and cars and stuff, and he built his own radio control system. I thought the whole process of engineering was fascinating.
Then came a movie that pointed that fascination in a new direction entirely.
Do you mind if I ask you a personal question?
I was struck by the computer HAL 9000 in 2001: A Space Odyssey. Open the pod bay doors, please, Hal. I'm sorry, Dave. I'm afraid I can't do that. What are you talking about, Hal? Even when I was a kid, I was fascinated by astronomy, mystery of like how human intelligence evolved. And 2001 was talking about all of those things. I thought this whole thing was incredible.
That curiosity took him from the suburbs of Paris to Bell Labs in New Jersey, perhaps the most storied research institution in the world, where he was given a data set of 9,000 handwritten digits and asked to teach a machine to read them.
The post office was interested in sorting mail, right? So, just reading the zip code. What I was interested in is, can you train a neural net to do visual perception? And so, I designed this architecture called convolutional neural network.
By 1996, American banks were using it to read millions of checks a day. But even as a field he helped to create exploded, LeCun grew more skeptical, not less.
So, the LLMs accumulate a lot of declarative knowledge and develop a little bit of understanding of the underlying structure of the world described by text. But their understanding is very shallow. They're mostly rote learners. Okay? They can regurgitate knowledge very well, and they can accumulate an enormous amount of knowledge, more than humans. But they're not very smart.
Not everyone who shared his Turing Award agrees.
You'll hear people saying things like, 'They're just doing auto complete. They're just trying to predict the next word.' But if you think about it, to predict the next word, you have to understand the sentences. You have to be really intelligent to predict the next word really accurately.
There's a joke that circulated recently. I want to wash my car and the car wash is 100 yards from my house, should I walk? And the thing says, 'Yeah, you should walk.' Okay, there is no common sense, right?
The solution, LeCun believes, is systems that learn not from text, but from direct experience of physical reality. He calls them world models, and he recently co-founded AMI Labs, a startup that's raised more than $1 billion to pursue his world models approach to general machine intelligence.
We need three or four like conceptual breakthroughs first of all to get to at least a viable human-level AI system. So, the scenario that, you know, a super intelligent system like takes over the world the next day is ridiculous. Perhaps the best analogy of the effect of AI on society would be similar to the effect of the printing press. It allowed the dissemination of knowledge in science and philosophy, and it made people smarter. Without the printing press, you wouldn't have the American Revolution and French Revolution, and we'll still be under the feudal system, right? Making people smarter is intrinsically good.
