humans& Details Off-Policy RL for Long-Horizon Rollouts, Trained on Its Own Multi-Thousand-GPU Cluster
gharik · x · 2026-10-21
humans& shared details of how it trained Ampersand:
- The team developed a new off-policy RL algorithm for extremely long-horizon rollouts and an efficient multi-agent training framework, scaling training across thousands of GPUs in its own cluster.
- Eric Zelikman argues owning hardware makes money go much further than renting, retains asset value after 3-5 years, and enables a differentiated model bet.
- DeepInfra, NVIDIA, and Supermicro helped build the cluster.
More from Infra
- Amazon and Google launch PyTorch tools to ease switching away from Nvidia chips — nordicinst · 2026-10-21
- Brev founding engineer Tyler Fong joins GPU cloud startup National Compute — blelbach · 2026-10-21
- OvertChat: a polished self-hosted ChatGPT-style UI for local models, one-command install — anitamaxwynnn69 · 2026-10-21
- Same B200, 50% More Output: SemiAnalysis on Why the Operator Matters — AI Engineer · 2026-10-21
- PyTorch once anchored NVIDIA's CUDA moat — the moat keeps changing — ivan_bezdomny · 2026-10-21
- Dev powers a Neural-MMO's AI brains with 12 AMD BC-250 mining cards, 3 nodes live — redfoxkiller · 2026-10-21