Yann LeCun’s AMI Labs Raises $1 Billion for Modular AI

Advanced Machine Intelligence Labs, founded by Yann LeCun after he left his role as Meta’s chief AI scientist late last year, has raised $1 billion to pursue research into modular, domain-specific artificial intelligence. The 12-person startup says it may not produce a saleable product for up to five years.

AMI Labs plans to build AI systems from smaller, specialized components rather than single, general-purpose language models. The proposed architecture includes a domain-specific world model, an actor that proposes actions using reinforcement learning, a critic that evaluates options against rules and short-term memory, a perception system tuned to relevant sensor data, short-term memory, and a configurator that routes information between modules.

Each instance of the architecture would be trained on directed data tied to a specific environment or role instead of broad internet text. AMI Labs says the relative importance of each module would vary by task-for example, a more comprehensive critic for applications that handle sensitive data or a stronger perception module for systems that must respond to video or audio inputs quickly.

The lab projects many modules will require only a few hundred million parameters and much less GPU power than today’s largest language models. AMI Labs argues that smaller, specialist models could run on-device or on much smaller cloud footprints and that the modular approach could reduce compute and data needs for defined tasks.

Investors committed the $1 billion despite the startup’s small team and long development timeline. The funding will support research into architectures, training methods, and engineering to run specialist modules efficiently. AMI Labs intends to focus on research work rather than immediate commercial deployment.

The lab cites previous machine-learning examples where focused systems mastered specific games or narrow tasks as evidence that specialized training can succeed. AMI Labs’ approach contrasts with generalist models trained on massive text corpora by major AI providers.

LeCun and his team say they will refine modular designs over several years; the company has not announced a product release schedule, named partners, or disclosed investor identities.