The First Fully General Computer Action Model
Source: https://si.inc/posts/fdm1/ Author: Standard Intelligence Team Date: 2026-02-24
Summary
Standard Intelligence announces FDM-1, a model trained on 11 million hours of video to predict computer actions. Claims it’s the first “fully general” computer action model — meaning it can operate arbitrary desktop software without task-specific training. Trained purely on observation (no RL), the model learns action sequences from watching humans use computers.
Key Claims
- Training data: 11M hours of computer-use video; no synthetic data, no RL — pure behavioral cloning at scale.
- “Fully general” means: can generalize to applications and UIs never seen in training, based on visual understanding.
- Key insight: computer interfaces share a small vocabulary (click, type, scroll) but infinite contexts — visual generalization is the hard part.
- FDM-1 outperforms existing computer-use models on GUI benchmark tasks.
- Implications for agents: if CUAs (computer use agents) can generalize, the software integration problem shrinks dramatically.
Entities
- Standard Intelligence — startup building computer action models
Concepts
- Coding Agents — adjacent; CUA is the “agent that uses software” version
- Autonomous Research — CUAs unlock many research automation tasks
- Synthetic Data — notable for not using it: behavioral cloning from human video