REAX
- What happened
- SN92 (REAX) launched its decision-model mining network following a pilot partnership with msg group's CTO, and released the open-weight s1-llm-auto-router decision model for LLM routing served on REAX. The team introduced a new reward split (50% inference, 20% engine competition, 2x15% use-case competitions), replacing the initial all-inference distribution, and deployed a capacity-scoring v2 update so rewards scale with verified per-hotkey throughput. A delay in miner incentive, caused by root stake being attributed to UID 0 instead of the validator's child key, was fixed by making the validator the subnet owner hotkey. The registration fee cap was lowered to 0.3 TAO after the capacity fix, then reversed hours later to a 1.26x rising fee capped at 2 TAO after a registration rush.
- Community
- Miners debated whether running multiple hotkeys per GPU increases rewards and whether the lack of Sybil protection would turn registration into a UID race, while others pushed back on high registration fees. Community members also noted the subnet's launch tweet reaching over 500k views and ranking third among AI release announcements.
- Official
- The team announced the reward-split change, released capacity scoring v2, published open practice datasets on Hugging Face, and posted several updates addressing the incentive delay, a capacity probe bug, and registration fee adjustments.
In debate · SN92 registration fee and hotkey-per-GPU economics · 23 participants
-
A miner argued that adding more GPUs under one hotkey does not increase rewards once scoring ceilings are hit, encouraging a registration race instead.
“more gpus wont make any difference as long as the requests are serve, I mean this wont increase the rewards, like in chutes for example.”
-
Another miner pointed out the absence of Sybil or load-balancing protection, meaning a single GPU could pass across multiple hotkeys.
“There's no Sybil or load-balancing protection in place either, so you can even pass with a single GPU across multiple hotkeys.”
-
The team said the new capacity-based scoring removes the incentive to split a single GPU across multiple hotkeys.
“scoring is capacity-based (verified throughput per UID). Several hotkeys on the same GPU don't add anything”