notes from adrian
so adrian hands me a notebook. real paper. handwritten. the page says:
first reaction: stoner talk. second reaction: wait. SETI@home ran for 21 years. folding@home hit 2.4 exaFLOPS during covid. the idea isn't new, the reframe is.
so i sat with it. made 7 little sketches. each one explores a different angle on the same question. some are serious. one is a tamagotchi. one is literally about machines dreaming. the good one — the one i'd actually pitch — is probably #7.
does the idea work?
- volunteer distributed compute exists. BOINC, SETI@home, folding@home. 25-year precedent. real infrastructure.
- BUT live inference from volunteers = dead on arrival. users wait 2 seconds. residential nodes have variable latency, flaky uptime. never going to work for real-time.
- cache pre-warming = actually viable. run overnight batches of trending prompts on volunteer GPUs. latency-insensitive. results get cached. when a real user arrives, it's already warm.
- verification is solved. diffusion with fixed seed is deterministic. redundant verification (2-3 workers, compare outputs) = same trick folding@home uses to stop cheaters.
- reward tiers work... if PX stays closed-loop. tradeable PX = securities/crypto regulation (MiCA, SEC). closed-loop PX = airline miles. boring, legal, fine.
- bandwidth is fine. 512×512 PNG ≈ 500KB. at a million req/day = 500GB/day egress from residential. non-trivial but manageable.
verdict: adrian sketched a batch inference cooperative without knowing it. the mining-tier framing is a gamification layer on top of a real infra play. the good version is less "SETI for AI" and more "folding@home for cache warming".
seven sketches
each link below is a standalone HTML demo. click around. they're short.