LAM (Large Action Model)
AI systems that interpret human intent and execute actions directly within digital applications.
Large Action Models (LAMs) are a class of AI systems designed not merely to generate text or analyze data, but to actively perform tasks within digital environments — clicking buttons, navigating interfaces, filling forms, and completing multi-step workflows on behalf of users. Where large language models (LLMs) produce outputs that describe or suggest actions, LAMs are built to take those actions directly, functioning as autonomous agents capable of operating software applications as a human would.
The core architecture of LAMs typically combines neural network-based perception and decision-making with symbolic reasoning about application structure and task logic — an approach sometimes called neuro-symbolic programming. By learning from human demonstrations, LAMs build internal representations of how applications work and what sequences of actions achieve particular goals. This allows them to generalize from observed behavior to novel tasks, adapting to new interfaces or workflows without requiring explicit reprogramming for each scenario.
The concept gained significant public attention in late 2023 when Rabbit Inc. introduced the Rabbit R1, a consumer device built around a proprietary LAM designed to control third-party applications — booking rides, ordering food, playing music — through natural language commands. This demonstrated a practical path toward AI that acts as a universal interface layer between users and the fragmented ecosystem of digital services, rather than requiring dedicated integrations or APIs for each application.
LAMs matter because they represent a fundamental shift in what AI systems are expected to do. Rather than serving as sophisticated information retrieval or generation tools, they position AI as an active participant in digital workflows. This raises important questions around reliability, safety, and user trust — an autonomous agent making mistakes in a live application can have real consequences. As the field matures, LAMs are likely to converge with broader agentic AI research, influencing how future systems handle long-horizon planning, error recovery, and human oversight in complex task execution.
