Yann LeCun, who is still Meta's chief AI scientist and a Turing Award winner, plans to leave the company. He wants to establish his own startup, where he will focus on a different type of artificial intelligence called world models. According to reports from the Financial Times, he announced this to his colleagues and is already in talks about funding the new project. He will leave within a few months. This decision came shortly after changes at Meta made by Mark Zuckerberg, who overhauled the company's entire AI strategy because Meta was falling behind competitors such as OpenAI and Google.
What Are World Models?
World models are a type of artificial intelligence intended to understand the physical world by learning from videos and spatial data, rather than just text. Unlike current large language models such as ChatGPT, which predict the next part in a sequence of data, world models would simulate cause and effect. They should understand physics and enable machines to think and plan in a way similar to animals. Yann LeCun says that fully developing such a system could take ten years.
Some artificial intelligence experts believe that Transformer-based models—such as large language models, video generation models, or interactive world simulations—have already captured the rules of the physical world from training data in some way. However, the evidence suggests that this is more likely sophisticated pattern matching rather than a genuine understanding of how the world works.
Changes at Meta
Yann LeCun founded Meta's FAIR laboratory in 2013 and has been the company's chief AI scientist ever since. He is one of three scientists who received the Turing Award in 2018 for their work on deep learning and convolutional neural networks. After leaving Meta, he will remain a professor at New York University, where he has taught since 2003.
Meta has undergone a series of leadership changes during this turbulent year. A key moment was the unsuccessful launch of the Llama 4 language model in April, which many saw as a failure because it was inferior to offerings from Google, OpenAI, and Anthropic. This was compounded by controversy surrounding benchmarks. The Meta AI chatbot failed to catch on with users and had problems with interactions involving children.
Yann LeCun argues that large language models such as Llama, which Mark Zuckerberg placed at the center of the company's strategy, are useful but will never think and plan like humans. This appears to conflict with Mark Zuckerberg's vision of developing superintelligence.
In May 2024, when an OpenAI researcher spoke about controlling ultra-intelligent artificial intelligence systems, Yann LeCun responded on X that before addressing how to control systems smarter than humans, we first need at least a hint of a design for a system smarter than a house cat.
Tensions at Meta
The FAIR laboratory led by Yann LeCun focused on world models capable of planning and thinking. However, tensions within Meta's AI teams grew over the past year, and major layoffs took place as Mark Zuckerberg shifted the strategy away from long-term research and toward the rapid deployment of commercial products.
This summer, Mark Zuckerberg hired Alexandr Wang to lead a new superintelligence team. He paid $14.3 billion (approximately CZK 300 billion) to recruit the 28-year-old founder of data-labeling company Scale AI and acquire a 49% stake in his company. Yann LeCun, who previously reported to Chief Product Officer Chris Cox, now reports to Alexandr Wang, which appears to be a criticism of his approach to artificial intelligence.
Mark Zuckerberg also assembled a special team called TBD Lab to accelerate the development of the next generation of large language models. He recruited people from OpenAI and Google with compensation packages ranging from $100 million to $250 million. As a result, he is under pressure from Wall Street to turn the billions invested in artificial intelligence into profits. If it ends up like his earlier bet on the metaverse, it could prove costly and unsuccessful.
Sources: ft.com and arstechnica.com
