- Sources: primary, discussion
- Summary: The 31-page report, submitted 2026-09-19, describes a production platform exposing FnCall, container, microVM and full-VM backends through one SDK, composing environments from independently versioned layers and loading image data on demand from the 3FS distributed filesystem. It gives operating figures rather than benchmarks: one production-scale unit spans roughly 160 nodes, serves about 3 million sandboxes per day, supports over 380,000 concurrent sandboxes and sustains over 5,000 sandbox creations per second. It also covers decoupling stateful rollout execution from preemptible GPU training, and mitigating reward hacking.
- Why it matters: It is the operator-side account of the same class of reinforcement-learning sandbox infrastructure that the day's OpenAI incident reports describe from the inside.
send feedback on this story