Yapay Zeka · 4 dk okuma · 3 Mayıs 2026
Synthetic Computers Enable Agent Training at Scale
Researchers create realistic digital workspaces to train AI agents on long-horizon productivity tasks, scaling from thousands to potentially billions of simulated user environments.
Researchers built a system to generate realistic computer environments and run month-long agent simulations to create training data for productivity AI.
- — Synthetic Computers at Scale generates realistic folder structures and content-rich documents to mimic actual work environments.
- — Agents simulate month-long productivity workflows, navigating filesystems, coordinating with collaborators, and producing professional deliverables.
- — 1,000 synthetic computers were created; each simulation ran 8+ hours with 2,000+ agent turns on average.
- — Training signals from these simulations improved agent performance on both familiar and unfamiliar productivity tasks.
- — Methodology scales to billions of synthetic user worlds with sufficient compute, covering diverse professions and contexts.
- — Approach positions synthetic environments as a foundation for agent self-improvement and reinforcement learning in real work scenarios.
Sık sorulanlar
- Synthetic computers are realistic digital environments with folder structures and documents that mimic actual workspaces. Agents simulate long-horizon tasks—like completing a month-long project—within these environments. The agent's successes and failures generate training signals that improve its ability to navigate real computer systems, coordinate with others, and produce professional outputs. This avoids relying on private user data or expensive human annotation.