Globally distributed pre-training across 48 A100s at $0.12 per million tokenstplr.ai·2 pts·synapz_org·0
PULSELoCo: 17x less trainer-to-trainer bandwidth in distributed RL post-trainingarxiv.org·3 pts·synapz_org·0
Autonomous RL Fine-Tuning on Ephemeral GPUs: Extending Karpathy's Autoresearchtemplarresearch.substack.com·5 pts·synapz_org·2
Pulse: Decentralized RL training centralized speed (100x weight sync reduction)arxiv.org·2 pts·synapz_org·0