Pluralis' decentralized training: model weights never fully materialize, no one owns the model
StewartalsopIII · x · 2026-10-25
Podcast host Stewart Alsop invited Alexander Long for an interview on decentralized AI, attaching prep notes detailing Pluralis Research's approach: model-parallel (not data-parallel) training splits models into pipeline stages across nodes so no single party ever holds full weights; the complete model exists only within the protocol; a Nesterov-based method fixes stale-update latency, demonstrated by training a GPT-1.3B over 200 Mb/s networks with only 2x slowdown; contributors earn partial model ownership enforced by the protocol. Research scientist Erfan Miah joined after "what Anthropic just did," arguing decentralized AI is the only way to keep AI good for humanity, bringing RL weight-sync work 100x more efficient.
More from Companies & People
- Ruff creator on moats: 'Caring is the moat' — visceral pain when things break — charliermarsh · 2026-10-25
- GDG Callao DevFest talk shows building a local Socratic tutor with Gemma — leslysandra · 2026-10-25
- Dario Amodei's high school teacher recalls he was genuinely worried about Y2K — menhguin · 2026-10-25
- Microsoft launches Agentic DevOps Engineer Expert certification via Exam AZ-401 — adnan_hashmi · 2026-10-25
- Figure CEO shows off three generations of its humanoid robots — adcock_brett · 2026-10-25
- Dario Amodei's Former Physics Teacher: He Was in the 'End of the World Camp' Over Y2K — Polymarket · 2026-10-25