banger decentralized async training idea req:
model parallelism + eigenlayer bs + eyeball scanning level sign ons of everyone's iphone
========>
everyone does 1 matmul op
do this and we get gpt5
found @HyperspaceAI / edgellama, seems interesting, basically tokens? as incentive mechanism to bootstrap inference clusters
they're prolly fronting the nodes in community servers that can run larger models though, advertising 4 petaflops seems disingenous without clear stats
@HyperspaceAI its main bottleneck for interesting stuff is it needs individuals with enough compute to run larger models to sign on.
registry of nodes seems pretty centralized as well for now, also questionable on whether it can support low latency inference reqs at large
@sheikhswampert That said these tokens are one of the few high beta proxies for betting on AI directly
Branding as anything other than a token/ instrument is not genuine
@sheikhswampert These things have incredible amounts of technical knowhows past a few guys in industry labs and based anons it’s useless for most and doesn’t need to be served at scale.
Self hosted software is evidence why this shits not necessary