Large action models move beyond chat to carry out tasks

Large action models (LAMs) are being developed to execute tasks rather than simply generate content. Trained on demonstrations and actions, they can interpret natural language requests and use software interfaces or function calls to complete activities such as booking travel, ordering groceries and sending messages.

Rabbit’s $199 R1 device, launched in January, was set to begin shipping within a week. Reports indicated at least 50,000 units had been pre-ordered in the two months since launch. Rabbit says its model learned from human interactions with apps including Spotify and Uber, allowing it to navigate interfaces and learn custom actions. Users grant access to supported services through Rabbit’s online portal rather than signing into apps on the device.

Microsoft is working on action and behavioral models for complex tasks across digital and physical environments. Salesforce is also exploring agents that combine language capabilities with independent decision-making, while emphasizing the need for stronger trust safeguards. Potential business applications include consumer banking, travel and patient management, though specific use cases remain emerging.

Shorthills AI co-founder Paramdeep Singh estimated that LAMs could take six months to two years to mature. Experts expect task automation to improve business returns, but warn that agents choosing services for consumers could disrupt competition between apps. Autonomous action also raises risks involving security, job displacement and other ethical challenges.