humans&ai Introduces GPU-Accelerated Weight Teleportation for RL Sync

Large-scale RL requires trainer and sampler weights to stay closely aligned, making terabyte-scale weight transfer a bottleneck. humans&ai proposes a 'weight teleportation' paradigm built on delta weight sync, with a Blackwell-native compression/decompression pipeline combining massively parallel GPU compression, distributed ownership with fault tolerance, a curated double-compression algorithm, and heterogeneous CPU-GPU decompression. The authors claim better-than-intra-cluster sync speeds in most cases, even when samplers and trainers are not co-located.