A top AI thinker is leaving Meta to start a new startup focused on Advanced Machine Intelligence (AMI), aiming to build world-model AI systems that understand the physical world, remember persistently, reason, and plan, with Meta remaining a partner.

The departure was announced on social media and comes amid Meta's leadership reshuffle and mixed results from its Llama family, signaling a shift in research direction.

Since joining Meta in 2013, he built and led Facebook AI Research (FAIR), which evolved into Meta’s flagship AI lab, and his move marks a dramatic pivot for the organization.

He argues that current emphasis on large language models diverts attention from other promising approaches, underscoring the need for diverse AI research paths.

He has long criticized language models for lacking core capabilities like physical-world understanding, stable memory, robust reasoning, and hierarchical planning.

Analysts expect significant longer-term effects across autonomous systems, healthcare, and education as AI capabilities advance beyond text-based models.

AMI will focus on autonomous learning and operating beyond pre-programmed instructions, with emphasis on robotics, perception, vision, NLP, and autonomous systems.

His philosophy centers on reconnecting perception, understanding, and reasoning, advocating for learning from sources beyond web-scale text rather than relying solely on massive language models.

AMI is envisioned as a broadly applicable system, potentially overlapping with Meta’s interests but designed as an independent venture to maximize societal impact.

The startup is positioned to pursue world-model AI that can imagine and predict outcomes in dynamic environments, such as predicting how a physical object would move, based on demonstrations in interviews.

The move signals a broader industry shift toward world-model and embodied AI versus scale-driven LLMs, with implications for the research landscape.

The departure intensifies the ongoing debate over whether AGI will emerge primarily from scaling LLMs or from alternative embodied and model-based approaches, highlighting a strategic and philosophical fork in AI research.