Yann LeCun has raised $1 billion for a 12-person startup built around the belief that large language models are not the route to broadly intelligent machines. Advanced Machine Intelligence Labs, or AMI Labs, plans to spend as long as five years on research before producing a commercial product.
LeCun left his position as Meta’s chief AI scientist late in 2025. His new company is pursuing modular, purpose-specific systems that learn how a domain works rather than relying mainly on predictions over internet text.
The funding reflects investor appetite for a credible alternative to increasingly large and expensive general models, even when the path to revenue is unusually long.
Intelligence is divided into specialised modules
AMI Labs’ proposed architecture begins with a world model representing the environment in which the system operates. An actor proposes actions using reinforcement learning, while a critic evaluates options against memory and hard-coded constraints.
A perception module handles the relevant input—video, audio, text or images. Short-term memory preserves current context, and a configurator orchestrates information among the components.
The balance can change by use case. A system handling sensitive information may need a stronger critic. A robot reacting to physical events may place greater weight on perception and a domain-specific world model.
Each module would train on information relevant to its purpose. That resembles successful systems that learn games through interaction rather than absorbing general language. It also contrasts with LLMs that use broad text training and prompt or reasoning loops to improve best-guess answers.
Smaller models could change AI economics
Frontier language models have consumed more resources with each generation. Reasoning approaches can add further inference cost by recursively processing intermediate steps before returning an answer.
LeCun’s modular systems could use a few hundred million parameters for specialised tasks rather than hundreds of billions in one general model. Components might run with a fraction of current GPU demand or directly on devices.
That would make AI cheaper, local and potentially more accurate within a defined environment. It could also improve privacy and latency by reducing dependence on remote infrastructure.
Modularity does not guarantee efficiency. Several components must communicate, remain consistent and learn from different data. Engineering and evaluation may become more complex even if each model is smaller.
The economic case therefore depends on whether targeted performance outweighs the convenience of one general system that can be prompted for many tasks.
A billion-dollar wager still needs evidence
LeCun’s criticism is that current LLMs cannot improve enough to fulfil the broad claims made for them. AMI Labs offers investors a different path while retaining a familiar startup promise: substantial future capability before commercial proof.
The five-year research horizon gives the team room to test foundational ideas, but it also delays market feedback. World models must demonstrate that they can represent complicated domains better than scaled language and multimodal systems already under rapid development.
Purpose-specific training raises practical questions. Who supplies the directed data, how easily can a system transfer to a new environment and what happens when the world changes beyond its model? Modular critics and hard-coded rules also need governance to avoid encoding incomplete assumptions.
AMI Labs’ small team and enormous funding make it an unusual experiment in concentrated research. The company does not need to win every AI task to validate its thesis. A clear advantage in robotics, industrial control or another domain where prediction over text is insufficient could establish a new architecture category.
The startup matters because it challenges the industry’s dominant direction at a moment when model costs and infrastructure demands are rising. Rather than make one general model larger, LeCun wants intelligence assembled from components that understand a particular world and can evaluate actions within it.
Investors have funded the possibility at extraordinary scale. The next test is scientific: whether modular world models can produce reliable learning, planning and perception that current systems cannot—and do so with the lower cost AMI Labs expects.
