
Building on bittensor - sn91 (@cascade_sn91)
Planum Temporale
Joined January 2025
- Tweets152
- Following244
- Followers239
- Likes465
Great work on video.affine.io
A week ago @cascade_sn91 moved from cold starts to warm starts.
Since then the competition has been producing checkpoints, and there are now enough of them to do something a single starting point can't.
The strongest models of each reign, kings and challengers both, form a pool. All of them scored on public benchmarks: CRPS and MASE across GIFT-Eval, BOOM and TIME.
Each round draws one init from that pool, and every generator trains from it. The controlled comparison is intact. What rotates is which member of the pool that init is.
A single init forever is one trajectory, chosen early. Rotating inits open a space of curricula through the same competition.
It changes what miners have to build, too. A generator tuned to one checkpoint only wins the rounds that draw its init. Data that helps from any starting point in the set wins every round.
Every round is signed and published: training records, scores, checkpoints.
Warm starts compound what the data teaches. Rotation makes sure it generalises.
Can't open the landing page for @Hippius_cloud registry (Bittensor's answer to Huggingface) without @cascade_sn91 models taking over the front page 😂







