humans& unveils weight teleportation for large-scale RL: sampler pause under 0.23s
ericzelikman · x · 2026-11-03
humans& released "weight teleportation," a technique that drastically speeds up trainer-to-sampler weight updates in large-scale RL training.
- Sampler pause time under 0.23s, with 3s non-blocking compression + prepare; the bottleneck becomes the GPU's decompression ASIC
- Achieved via massively parallel GPU-based compression of weight deltas, a heterogeneous decompression and patching pipeline co-designed with hardware-native features
- Enables samplers anywhere ("maybe even space!"); built on and contributed back to the Miles framework's delta weight sync
More from Infra
- SEC filing shows Anthropic committed up to $120.5B in compute spend through 2029 — arthurcolle · 2026-11-03
- Unverified: SEC filing claims Anthropic committed up to $120.5B through 2029, ~$3.1B/month for xAI infrastructure — arthurcolle · 2026-11-03
- Ubuntu's delayed CVE-2026-98432 fix is in -proposed, likely held back by database regression fixes — bytebot · 2026-11-03
- The Circuit: Anthropic's S-1, Intel's questions, Marvell and custom silicon bets — BenBajarin · 2026-11-03
- Analyst: Palantir earnings could hurt the AI trade more than help it, squeezing rivals alongside Nvidia — pdamodaran · 2026-11-03
- Inference providers can't survive without DeepSeek-grade KV-cache engineering, argues thread — AccBalanced · 2026-11-03