Bittensor’s (alleged) $10M rug pull (feat. Mark Jeffrey) | E2275

What It Covers

Stillcore Capital Partner Mark Jeffrey joins This Week in Startups to break down the alleged $10M rug pull on Bittensor's Templar subnet by founder Sam Dare, who dumped 37,000 TAO tokens on subnet holders before departing. The episode also features three subnet founders from Bitmind and Macrocosmos explaining their deepfake detection and distributed AI training projects.

Key Questions Answered

• •Subnet Rug Pull Mechanics: A subnet owner accumulates tokens automatically through chain emissions. If they hold a majority of subnet tokens, they can flood the on-chain liquidity pool — essentially a decentralized exchange — converting those tokens to TAO, then selling on Binance for cash. Sam Dare allegedly executed this with 37,000 TAO worth roughly $10M, causing TAO's price to drop approximately 25% immediately after.
• •Proposed Fix — Conviction Locking: Konstantin Steves released a governance proposal where subnet ownership is determined by who locks the most tokens for the longest period. Active subnet operators would have tokens locked while running the subnet, preventing sudden mass dumps. This mechanism rewards capital conviction and makes hostile exits structurally difficult, replacing trust-based systems with smart contract enforcement.
• •Proper Subnet Exit Protocol: If a subnet owner has grievances, the correct path is negotiating directly with the foundation — agreeing on a fair token allocation for contributions to date, then transferring the subnet's private key to the foundation for redistribution. This preserves token holder value and continuity of the project rather than unilaterally liquidating positions and abandoning operations.
• •Bitmind's Adversarial Detection Model: Bitmind (subnet 34) runs two competing miner classes simultaneously — one optimizing deepfake detection accuracy, one generating synthetic content to fool detectors. This red-team structure forces continuous model retraining on a daily or weekly cadence, keeping detection ahead of new generative models without waiting for centralized dataset updates. Enterprise use cases include media verification, secure hiring, and legal evidence authentication.
• •IOTA's Interruptible Distributed Training: Macrocosmos subnet 9 (IOTA) enables frontier model training using compute units as small as a MacBook for as little as 20 minutes. By designing for interruptibility — assuming nodes drop out — they can source GPU compute at roughly 10 cents on the dollar from providers with idle capacity. The train-at-home app requires two clicks to install and pays miners in IOTA tokens for overnight compute contribution.
• •TAO vs. Subnet Token Investment Strategy: Buying TAO provides mutual-fund-style exposure across all 128+ subnets. Buying individual subnet tokens is equivalent to angel investing in a specific project — higher risk, higher potential return, requires active evaluation. Stillcore Capital currently holds approximately 80% in subnets and 20% in TAO, with plans to rebalance toward 70/30 subnets-to-TAO. Subnet tokens are liquid at any time, unlike traditional startup equity.