This week the Lunch with the FT is with Yann LeCun. It’s an interesting conversation in a (very expensive) Parisian restaurant.

I didn’t know that LeCun worked at AT&T Bell Labs in the 1980s and 1990s where he had access to the resources to play with convolutional neural networks.

In 2013 he became Meta’s Chief AI scientist, apparently under the conditions that he could retain his academic position at NYU and work from there and that all the research at FAIR is made publicly available. According to him he also was the main person to push for the open-weight strategy of the Llama models.

They also chat about the reasons why he left Meta in 2025. It sounds like a typical research vs. product story. He wanted to move away from the LLM-approach whereas Zuckerberg wanted to focus on their Llama models which led to the team rushing out the disappointing Llama 4 model. After that a lot of researchers left FAIR and Meta aqui-hired Alexandr Wang(Scale AI) as the Chief AI officer. Leading the new Meta Superintelligence Labs he continue to focus on productizable LLM-based models.

Meanwhile LeCun reemphasizes that LLMs won’t be able to achieve superintelligence because text as training data is too limiting for their learning. He believes that architectures like V-JEPA, called world models, are much more likely to achieve superintelligence. Those models are learning from the physical world via video and spatial data. He thinks that the new fundamental research labs, Mira Murati’s Thinking Machines and Sutskever’s Safe Superintelligence, are a better way to work on these new architectures than the frontier model companies.