How can decentralized AI win? It will never be as efficient as the centralized alternatives (because of communication overhead). One possibility is by aggregating otherwise underutilized hardware.
Comments
Log in with your Bluesky account to leave a comment
I am more intrigued though by the possibility to incentivize global human-machine collaboration on content, models, queries, etc. This has the potential for huge efficiency gains over centralized competing entities which duplicate all efforts.
Did you see this article in the Economist last week?
I'm far from an expert on the pros and cons of different training designs, but at a glance decentralized (well, distributed) training offers quality improvements not in spite of, but *because of* inefficiencies.
It's worth noting that Vincent Weisser is one of the OGs in DeSci (Molecule ecosystem), with a deep background in the design principles of decentralization.
Comments
I'm far from an expert on the pros and cons of different training designs, but at a glance decentralized (well, distributed) training offers quality improvements not in spite of, but *because of* inefficiencies.
https://www.economist.com/science-and-technology/2025/01/08/training-ai-models-might-not-need-enormous-data-centres
https://x.com/vincentweisser