AI pioneer Yann LeCun, fresh off a decade at Meta, is building a new class of systems at Paris-based AMI Labs aimed at tackling messy, real-world tasks that stump today’s large language models. Backed by more than $1 billion in seed funding from investors including Nvidia and Jeff Bezos’s family office, the startup is developing a Joint Embedding Predictive Architecture to reason about physical outcomes rather than predict text. LeCun argues LLMs are ill-suited for robotics and won’t scale to superhuman intelligence, a view echoed by researchers advancing “world models” that can simulate and explain cause-and-effect. AMI plans industrial pilots next year, betting that more grounded AI will unlock practical autonomy before general-purpose systems arrive.
Related articles:
— World Models
— Dream to Control: Learning Behaviors by Latent Imagination
— DreamerV2: Mastering Atari with Discrete World Models
— A Generalist Agent
