OpenAI's 10,000-agent, 130B-token run pushed slime v0.4.0 to rethink RL infrastructure scale
teortaxesTex · x · 2026-10-09
slime v0.4.0 cites OpenAI's Navier-Stokes effort—10,000 concurrent agents and 130B output tokens—as motivation for scaling RL infrastructure. Four architectural changes: sync training for algorithmic exploration plus fully async training for throughput; distributed rollout orchestration across cluster CPUs; a new storage middleware called straw with durable queues and shared tensor storage; and independent recovery that restarts Megatron while healthy SGLang engines keep running, preserving accepted rollout work.
More from coding & agent
- Open-Voyager: open-source creative agent billed as the Codex for video and design — matchaman11 · 2026-10-09
- Give AI Agents an Empty 3D World: They Build Villages and Even Write Constitutions — pkmital · 2026-10-09
- Top Chinese agent team studies codex and claude code to solve context management — aigclink · 2026-10-09
- Multi-agent bots built a 24/7 AI news TV station, and it's going open source — vista8 · 2026-10-09
- Agent Arena Leaderboard: Claude Opus 5.5 Tops GPT 6 Astra Across 2.3M Agent Sessions — arena · 2026-10-09
- anti-slop: an open-source rulebook with 5.1k stars to strip generic AI slop from coding agents — tom_doerr · 2026-10-09