Citizen Distributed LLM Training PoC: Port 443 Only, Safetensors, MultiKrum
OrganizationTop9026 · reddit · 2026-11-14
The author connected their HiveMind coordinator with CrowdGPT (led by Vxtzq) for a citizen-trained distributed run around Crowd-v1, an open base model with a 658M transformer trunk (1.01B total params with a 151k-vocab CrowdEngram table), targeting Nov 14–15 CET with a five-minute setup. Addressing why 8–16GB GPU owners usually sit out distributed runs: - communication uses outbound-only HTTPS on port 443, zero inbound listeners- weights move strictly as SHA-256-verified safetensors, no torch.load/pickle- updates pass MultiKrum outlier filtering plus held-out loss checks before merging via DiLoCoThe author candidly notes this is an experimental PoC, not a polished cloud service — don't send private prompts or sensitive keys through volunteer nodes.
More from Infra
- AMD's Rumored Siren Halo APU Reportedly Packs Dedicated GDDR7 On-Package Memory, Eyeing Late 2028 — Terminator857 · 2026-11-15
- GLM-5.3 on Two DGX Stations Peaks at 607.8 tok/s but 32 Concurrency Causes 453s Latency — petrusenko_max · 2026-11-15
- Pundit likens possible Huang exit to Microsoft antitrust moment as rivals circle Nvidia — TiernanRayTech · 2026-11-15
- Analyst warns of hyperscaler monopoly: NVIDIA loses pricing power, startup funding dries up — BenBajarin · 2026-11-15
- Anthropic CEO Dario Amodei explains why the company buys compute conservatively — _AustinCalvert_ · 2026-11-15
- Anthropic IPO deck chart sparks debate: Mythos could be retrained with a fraction of the compute — Witty_County5128 · 2026-11-15