Bittensor creator Const on Affine, dTAO, "mining reasoning," and more | E2326

What It Covers

Jacob Steeves, creator of Bittensor, explains how the network applies Bitcoin's permissionless proof-of-work model to artificial intelligence. The episode covers subnet architecture, the dynamic TAO (dTAO) economic system, the Affine subnet's "mining reasoning" approach, and why decentralized AI represents a viable third path against OpenAI and frontier labs.

Key Questions Answered

• •Permissionless market design: Bittensor's core competitive advantage mirrors Bitcoin's: anyone anywhere can contribute compute anonymously without approval. This creates hypercompetitive markets that scale efficiently — when one subnet tripled miner payments, compute supply tripled within two months. Human organizations cannot replicate this because hiring, contracts, and coordination create friction that permissionless code-based mechanisms eliminate entirely.
• •Subnet economics (dTAO): Each of Bittensor's 128 subnets issues its own alpha token paired to TAO via an automated market maker liquidity pool. TAO has a 21 million cap identical to Bitcoin, with one token minted roughly every 12 seconds post-halving. Subnet registration costs approximately 600 TAO (~$121,000), set via Dutch auction that doubles on registration then declines until the next buyer.
• •Adversarial design is the entire product: Steeves states that resisting bad actors is not a feature of subnet design — it is the entirety of the project. Subnet 51 built GPU attestation without trusted execution environments by SSH-verifying hardware remotely. Subnet ENGY uses cryptographic proofs (TopALock, developed by Prime Intellect) to verify that miners are running the correct model at inference time, not a cheaper substitute.
• •Decentralized training bandwidth problem: Training large models requires merging weight gradients every step, generating terabytes of data transfer per iteration. Across hundreds of thousands of training steps, bandwidth requirements exceed what distributed home or cloud nodes can sustain. Bittensor's Templar subnet is attempting to solve this via S3-compatible bucket storage as a shared synchronization layer, which would unlock trillion-parameter decentralized training.
• •Affine subnet — mining reasoning: Affine (subnet 120) targets reasoning capabilities rather than raw model size. Miners produce small models or adapters that improve how larger models think through problems, enabling faster inference — potentially 30x — by offloading reasoning to a specialized lightweight model. The long-term goal is using mined reasoning data as training signal to produce frontier-competitive models through the permissionless network.
• •Rug pull mitigation via voluntary locking: After the Templar subnet collapse — where founders sold holdings and publicly blamed the network — Bittensor introduced optional token-locking mechanisms for subnet teams. Teams can publicly commit to multi-year locks, making large sell events visible on-chain before execution. This mirrors startup vesting structures and gives investors an exit window when founders signal intent to liquidate.