Humans& shares weight teleportation: <0.23s sampler pauses for large-scale RL, bottleneck shifts to GPU ASIC

niloofar_mire · x · 2026-11-03

Humans& unveiled "weight teleportation", attacking the slow, expensive trainer-to-sampler weight updates in large-scale RL.

Sampler pause time under 0.23s with 3s non-blocking compression + prepare — faster than existing intra-cluster techniques and at internet speeds, so samplers can run anywhere ("maybe even space!").
The approach uses GPUs for massively parallel compression of weight deltas, plus a heterogeneous decompression and patching pipeline co-designed with hardware-native features.
It's so efficient the bottleneck ends up being the GPU's decompression ASIC.

The upshot: RL samplers no longer need to live in the same cluster as trainers, loosening a key infrastructure constraint for large-scale RL.

More from Infra