Could a Folding@home-style P2P network of idle GPUs collectively run and train LLMs?
lenicalicious · reddit · 2026-11-03
A Reddit user proposes an open question: build a Folding@home-style decentralized peer-to-peer network that pools the world's idle compute — gaming PCs, servers, phones — to collectively run and even train large LLMs.
The vision: no central servers, subscriptions, crypto tokens, or corporate ownership; contributors donate spare resources in exchange for access; latency is acceptable if responses are slower; anonymity is architectural (no accounts, no IP logging, no centralized histories). The author acknowledges prior art like Petals and known obstacles — synchronization, malicious nodes, computation verification — but argues these may be underexplored engineering problems rather than fundamental limits. They invite distributed-systems and LLM-architecture experts to weigh in.
More from AGI Musings
- As AI Floods Software with Code, Products Spark Wonder but Rarely Delight — manosaie · 2026-11-03
- Agentic payments winners will create value, not build payment products, argues exec — schwentker · 2026-11-03
- Frontier model training is a monopoly business: the bull case for Anthropic's 'One Ring' thesis — abhiadesai · 2026-11-03
- 'Text-to-neural-network' vision: describe a task and get a purpose-built tiny net — flowersslop · 2026-11-03
- Micro schools will adopt AI tutors first and deliver '10x better' education, advocates say — RachelVT42 · 2026-11-03
- Yoav Goldberg: the 'of course AI can do that' bucket is getting insane — yoavgo · 2026-11-03