The godfather of AI who argues the field has lost its way

(Representative Image)

Yann LeCun has spent four decades shaping the field of artificial intelligence. His research underpins many of the systems that now dominate the industry. But even as the rest of the tech world races to build ever larger language models, LeCun is convinced that the field has taken a wrong turn. The 65-year-old scientist has been increasingly out of step with Meta, the company that recruited him more than a decade ago. Recent reports suggest he may soon leave to pursue his own vision, focused on what he calls world models rather than the predictive engines that power today’s chatbots, the Wall Street Journal reported.

A disagreement at the top of Meta

Mark Zuckerberg has made superintelligence the centrepiece of Meta’s mission. The company has poured billions of dollars into building Llama, aiming to outpace competitors like ChatGPT and Google’s Gemini. LeCun has chosen another path. He has said repeatedly that today’s models cannot reach human-level reasoning and that scaling them further will not change that. He compares current systems to animals that react to patterns rather than understanding the world around them.

Several years ago, he stepped back from leading Meta’s AI research lab, FAIR, and became an individual contributor focused on long-term theoretical work. He now argues that within five years, architectures based on world models will make today’s large language models obsolete. At a recent symposium, he suggested that no serious researcher would rely on current LLMs once better alternatives appear.

The idea he wants to build

A world model is designed to learn by observing the physical environment, similar to how a child or animal forms an understanding of cause and effect. Instead of predicting the next word in a sentence, such a system would develop an internal representation of the world and how it behaves. According to people familiar with his thinking, LeCun has been discussing plans for a new company that would build such models from the ground up.

Meta declined to comment on the possibility of his departure, and LeCun did not respond publicly. But the interest in his research direction has grown as he speaks more openly about his disagreements with the dominant approach.

A long career of early insight

LeCun grew up in the suburbs of Paris and studied at the Sorbonne in the 1980s. He pursued machine learning before the field had a name, joining Geoffrey Hinton’s lab in Toronto before Hinton’s work became widely recognised. He later spent years at Bell Labs in New Jersey, where he developed handwriting-recognition systems still used by banks and worked on early tools for digitising documents.

He has said he learned most not from formal computer science training but from reading physics textbooks and discussing problems with researchers from different disciplines. His broad curiosity helped shape the techniques that decades later enabled modern deep learning.

In the early 2000s, he joined New York University and eventually founded its Center for Data Science. Zuckerberg recruited him personally in 2013 to lead Facebook’s new AI initiative. In 2018, LeCun, Hinton and Yoshua Bengio won the Turing Award for their breakthroughs in neural networks.

A changing place inside Meta

In recent years, his role inside Meta has become more symbolic. He was not involved in building the first Llama models and does not participate in day-to-day AI operations. Instead, he travels frequently, gives talks at conferences and continues to explore ideas that may not translate into products anytime soon.

Meta’s re organisation this summer brought further changes. Alexandr Wang, a 28-year-old entrepreneur who founded Scale AI, became Meta’s new chief AI officer. Shengjia Zhao, one of the creators of ChatGPT, became the company’s chief scientist. Internally, employees wondered what the restructuring meant for LeCun’s future. Some saw the shift as a quiet sidelining of a pioneer who disagreed with the company’s direction.

Zuckerberg insisted publicly that LeCun’s role was unchanged, calling him the chief scientist of FAIR. LeCun responded that he looked forward to working with Zhao. Still, employees say FAIR now has fewer resources and less influence as Meta pours money into practical, product-driven AI research.

A sharp critic of today’s dominant models

LeCun has not softened his views. In interviews and public talks, he argues that language models cannot achieve human-level intelligence because they lack grounding in the real world. He says they are good at producing fluent text but cannot understand or reason in ways that match even simple animals.

He tells young researchers to avoid working on LLMs if they want to push the field forward. He believes that the industry’s focus on scale has blinded it to more fundamental questions about perception, memory and learning.

A future that may diverge sharply from today

As he contemplates his next move, LeCun remains one of the field’s most influential voices. His ideas helped create deep learning, and he is convinced that the next leap in AI will come from a different set of principles. Whether his predictions prove correct will determine whether he is once again ahead of the curve or an icon standing apart from the mainstream.

For now, he continues to travel, debate, lecture and warn that today’s language-model boom will not last. And he continues to search for the architecture he believes will define the next era of artificial intelligence.