Bittensor subnet claims a frontier-matching model emerged from leaderless relay competition
bittingthembits · x · 2026-09-29
A thread on the Affine subnet in the Bittensor ($TAO) ecosystem: an open, leaderless relay race where anyone can improve the reigning champion model and submit it; if it wins on held-out tests it becomes champion, weights are published openly, and the winning miner gets paid.
Charts show scores climbing reign after reign — one miner spotted reign 15 was a dead end, rolled back to reign 13, and jumped to 18. The author stresses this strategy wasn't designed by anyone; it emerged from adversarial strangers coordinating with no boss.
The claim: this process produced a model matching a 2025 frontier model from Anthropic. Caveat: this is a community self-claim, not independently benchmarked.
Related event: Decentralized Bittensor Subnet Trains Model Rivaling Frontier AI(2 posts)→
More from Models
- Is Sonnet 5.5 as Good as Sonnet 3.5 Was? Users Ask if Anthropic Is Back — capetorch · 2026-09-29
- 'They never stopped training — just stopped testing on the live web': commentary sparks debate — dejanseo · 2026-09-29
- China's 7-day holiday means no DeepSeek release until Wednesday at best — teortaxesTex · 2026-09-29
- Shanghai AI Lab releases Intern Decision: 0.8B-4B Apache 2.0 models for structured decisions — AdinaYakup · 2026-09-29
- OpenAI reopens $200 Pro plan but halves API credits per dollar, pushing usage-based billing — The Decoder · 2026-09-29
- Building a Polish Legal Agent: Is a 50B+ Low-Active-Param MoE Worth It Over a 27B Dense Model? — SignificantZebra5883 · 2026-09-29