Reflection AI runs 1.3B sandboxes in largest RL run, peaking at 883K vCPUs
zacharynado · x · 2026-10-28
Reflection AI shared details of its largest RL run for Beam: roughly 1.3 billion sandboxes spanning RL training, evaluation, and ablations.
With evals and other RL experiments running, the fleet peaked at 883,006 vCPUs (83% of allocatable) across 23 clusters on Google Cloud and AWS.
The team says it will adopt OpenEnv, now part of the Agentic AI Foundation, for future runs, bringing its learnings into the open ecosystem.
More from coding & agent
- AI-native macOS Dock experiment sparks call for agent-friendly open-source distribution — trq212 · 2026-10-28
- CC Switch tutorial: rotating multiple Claude official subscriptions and ban-risk notes — dotey · 2026-10-28
- MCP 2.0 spec extended: servers can now publish subscribable events — ycombinator · 2026-10-28
- Ruff author: 'time to recompile' is a lossy metric when judging agent-driven Rust refactors — charliermarsh · 2026-10-28
- Thinker launches on Product Hunt: context compression and knowledge caching for coding agents — yoyoyoav · 2026-10-28
- A 4-step AI research workflow: GPT drafts, Gemini fact-checks, NotebookLM podcasts it — sanderssays · 2026-10-28