A new method developed by Humansand reduces the sampler pause time for reinforcement learning (RL) to less than 0.23 seconds. The software, known as “weight teleportation,” uses GPUs to perform parallel compression of model weight updates and integrates with Nvidia’s Blackwell chip architecture to allow these updates to be transmitted between clusters at internet speeds.
The system handles non-blocking compression and preparation in approximately 3 seconds, addressing a technical bottleneck that previously required training and sampling components to be located within the same data center to manage high data volumes. Humansand has released the technology as open-source to facilitate the deployment of decentralized RL across distributed hardware.
