Yann LeCun, who serves as Meta’s top AI scientist, has decided to part ways with the company. In a recent social media announcement, LeCun shared his plans to establish a new AI startup. This development coincides with a period of turbulence in Meta’s AI division.

A Meta representative disclosed to Business Insider that Yann LeCun, a highly respected figure in the field of AI, will be departing from Meta to launch a new AI project. LeCun made the news public on Facebook, expressing his intention to create a startup centered around his enduring interest in world-model research.

While Meta will collaborate with LeCun on his upcoming venture, specific details about this partnership haven’t been disclosed yet. The departure of LeCun comes at a time when Meta’s AI unit is experiencing internal challenges.

In recent months, Meta has onboarded numerous leading researchers and engineers from competitors and restructured its AI initiatives under the newly established Superintelligence Labs division, headed by former Scale AI CEO Alexandr Wang. Reports from Business Insider suggest that tensions have arisen among team members due to differing compensations and backgrounds, leading some existing researchers to consider leaving the company.

In August, Meta underwent a major reorganization of its AI operations by forming four distinct teams dedicated to research, training, products, and infrastructure. This strategic shift reflects Meta’s ambition to outshine competitors like OpenAI, Google DeepMind, and Anthropic in developing large-scale AI models.

Concurrently, there have been fluctuations in key roles among researchers within Meta, and the release of Meta’s Llama 4 received lukewarm responses internally and externally. Recently, Souminth Chintala, the creator of Meta’s open-source PyTorch AI framework, departed from Meta after an 11-year tenure to join Mira Murati’s Thinking Machines Lab.

LeCun’s departure doesn’t come as a complete surprise given his public criticism of an overreliance on extensive language models. He advocates for the JEPA approach, which focuses on training AI to comprehend and predict real-world scenarios from visual and sensory data rather than generating text. In contrast, Meta has been emphasizing the expansion of large language models (LLMs) and pursuing commercially driven model development.

If you have any information to share, you can reach out to Pranav Dixit via email at pranavdixit@protonmail.com or Signal at 1-408-905-9124. To ensure secure communication, use your personal email address, a non-work WiFi network, and a non-work device; refer to our guide for sharing information securely.