humans& announces ‘weight teleportation’ for reinforcement learning

humans& announced that it had open-sourced weight teleportation, a method for moving reinforcement-learning model updates from trainers to samplers. The design uses GPU compression and a CPU-GPU decompression pipeline. In its production RL configuration, the company reports sampler pauses below 0.23 seconds and end-to-end synchronization below four seconds. Those figures describe its setup, not a general speed guarantee.