01 / DAILY
Robots, phones, and cheap training runs
Opentensor's Novelty Search series turned to SN80, OpenRoboto, an open competition for robot intelligence where miners improve shared base models and each winning champion becomes the new baseline. The SN80 team is now pushing into real robot validation with a project called 'Shift'. Separately, a Chutes AI presenter at the Exploit Summit described training an 8B parameter model across distributed gaming GPUs for about $6,500 in rental costs — the resulting model reportedly runs entirely on a phone's CPU at close to 60 tokens per second.
02 / DAILY
SN81's compute-fleet mechanism
A contributor laid out the Reliquary team's design for SN81, where miners build compute fleets that search for training rollouts — a setup compared to how Conjecture's '66' project searches math proofs. The design relies on gamma tokens, which let the subnet scale by directing a portion of miner emissions toward building out compute fleets on one of Bittensor's compute subnets.
03 / DAILY
Economics debate resurfaces
One community member pushed back on ongoing protocol development, arguing the project should fix its core economic model before adding new features. Another took the opposite view, saying Gamma's composability could let subnets act as building blocks rather than isolated products, enabling AI services to buy from and build on each other. Separately, a community member said liquidity pool usage had prevented most recent subnet deregistrations.
