How Paper Instruments uses E2B to cut idle GPU time for RL rollouts

Paper Instruments develops frontier AI models for knowledge-work fields including consulting, finance, investment banking, and law, using benchmarks such as DRACO and DiligenceBench to train and evaluate agents on complex research tasks. Its reinforcement-learning workflows require thousands of simultaneous, fully isolated rollouts with separate filesystems and persistent states to prevent cross-contamination and reward hacking while maintaining traceable links between generated data, benchmark grades, and model training. After evaluating sandbox options and the alternative of building infrastructure internally, the company chose E2B for its faster command execution, resilience under production spikes, and managed sandbox capabilities. By creating initialized sandbox templates and snapshots that preserve filesystem and memory states, Paper Instruments can launch large numbers of rollouts from identical environments, reducing cold-start delays, idle GPU time, and training costs. The company also cites E2B’s documentation, debugging visibility, and developer experience as factors that help its team focus on improving training data and methods as it expands large-scale RL experiments and evaluation.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.