A Fraud, a Major Upgrade, and a New Direction for Bittensor

Mark Jeffrey joined Jacob ‘Const’ Steeves for a wide-ranging Novelty Search covering one of the most eventful stretches in Bittensor’s recent history.

The conversation opened with the Quasar ‘fraud,’ where the community caught an 18-year-old founding team copying roughly 95% of its weights from Ling before doubling down when confronted.

The same week also brought several meaningful milestones, including Engy’s inference subnet running Kimi K3 at world-cheapest pricing and Nous Research adding native Actual (SN95) support inside Hermes.

Root Reborn also delivered its first live results, giving the community an early look at how the protocol’s new incentive structure is working in practice.

The Biggest TAO News This Week

1. The Quasar fraud was caught by community analysis: A community member named Cass tipped off 95% weight copying from a Ling model.

When confronted, the team resubmitted a “corrected” version using sign-flipping obfuscation. Const bought them out of their contract and is now finding a new team to take the subnet 24 forward.

2. Engy built the cheapest working inference network in under a month: The team behind Engy shipped a functioning decentralized inference subnet running Kimi K3 at the lowest per-token rates on the market, now targeting a top-slot subnet position.

3. Nous Research added native Actual (SN95) support inside Hermes: Hermes Agent users can point their harness at personal Actual-powered compute directly through settings. Const called it a possible reunion moment for Bittensor and Nous.

Firing Up Hermes Agent on Your Own Compute with Actual (SN95)

4. Isabella Liu’s Connito (SN102) launched as Bittensor’s answer to Tinker: The subnet builds custom LoRA adapters for GLM at high speed, producing fine-tuned versions of large models on demand.

5. OpenRoboto is Bittensor’s first Shenzhen-based team: The team launched a robotics foundation model subnet from Shenzhen, building pre-trained models designed for fine-tuning into any specific robotics problem.

OpenRoboto (SN80) and the Competitive Market for Robotics Models

6. Parallax public release from Jon Durbin is imminent: Strong results on a 5B model, not yet incentivized, but the community consensus is he is working on something genius.

Parallax Made It Possible to Train Giant AI Models on Scattered, Ordinary GPUs

7. Root Reborn cut roughly 900 $TAO per day of automatic sell pressure: The old mechanism forced validators to sell subnet dividends to pay root stakers. Removing it moves the chain from a negative position on the network to a neutral one.

Top 10 Frequently Asked Questions About the Root Reborn Proposal

8. Root staker yield trajectory (9% down to 4% back up to 6%, with 15-25% projected): Const’s preferred way to think about it is not as yield but as allocation, giving root stakers exposure to 128 functioning projects rather than fixed-income equivalents.

9. Liquidity pool upgrade is the next major roadmap item: Current Balancer-style pools bound prices at both ends, meaning teams like Chutes (SN64), Lium (SN51), and Targon (SN4) literally cannot move much without full market cap injection. Derivatives, spot books, and on-chain order books come next.

10. Spec 444 may remove the 1-minus-miner-burn from emission: Lium (SN51) tripled its miner payout and revenue in two months at 100% emission. GM (SN28) and Hippius (SN75) expanded aggressively at the same rate. The team is now weighing whether to roll it back.

Const’s Plan of Handling Over Bittensor

Const’s current stretch of active intervention in Bittensor is explicitly temporary, with the goal of eventually handing over the keys. Absolute power builds karma even when used correctly, because people can start expecting intervention in every case.

Long-term, Bittensor needs self-organization from the ground up rather than centralized enforcement from the top. The path toward full permissionlessness runs through specific protocol upgrades and ultimately the removal of central oversight.