Mining BTC on Bittensor SN14 :: TaoHash :: Latent Holdings (+SN5 HONE Announced)

This episode reveals the strategic pivot of TaoHash's Bitcoin mining subnet and introduces HONE, a new venture aiming to solve AGI by pioneering decentralized hierarchical AI training on Bittensor.

Introduction: The Core Contributors Behind Bittensor's Infrastructure

The episode begins with an introduction to key figures from Latent Holdings, a team instrumental in maintaining and decentralizing the Bittensor ecosystem. Host Jake highlights the critical, often behind-the-scenes, work of developers like Cameron (aka "Vune"), who has been a core contributor since 2021, maintaining the Bittensor SDK, API, and other vital infrastructure.

Joseph, representing Latent Holdings, frames their mission as being synonymous with Bittensor's: to make the protocol accessible, understandable, and stable. Latent Holdings has taken over stewardship of essential open-source tooling, including the SDK, CLI, and developer documentation.

• Key Contributions: In just six months, the small team at Latent has produced over 100 releases, nearly 4,000 code commits, and changed over 173,000 lines of code, making them the largest open-source contributor outside the OpenTensor Foundation.
• Projects: They are the creators of Tao App (an explorer), TaoHash (Subnet 14), and support Learn Bittensor, a resource for demystifying the protocol.

TaoHash (Subnet 14): Decentralizing Bitcoin Mining

Abe, the lead for TaoHash, presents Subnet 14 as a category-defining project: an openly owned and accessible Bitcoin mining pool. The core problem TaoHash addresses is the centralization of Bitcoin's network, where just three pools control over 56% of the total hash rate.

TaoHash operates by directing miners' hash rate to the Bitcoin network, rewarding them with both the mined Bitcoin and TAO (the native token of Bittensor, here referred to as "alpha"). The subnet retains a small pool fee, which is returned to TAO holders who have staked on the subnet.

Abe explains the value proposition: "Miners pull hash rate and get paid in alpha. Mined Bitcoin is sent to the miners based on their shares and the pool fee is kept by the subnet. The pool fee is given to the holders returning them value on their stake."

• Hash Rate: This is the total combined computational power being used to mine and process transactions on a Proof-of-Work blockchain like Bitcoin.
• Scalability: Abe presents a formula showing that TaoHash can theoretically incentivize up to 7% of the entire Bitcoin network's hash rate at a $7 TAO price, demonstrating significant scaling potential.

The Original TaoHash Design and Its Flaws

The conversation, prompted by Jake, shifts to the "drama" around TaoHash's initial design and the critical lessons learned. Abe candidly breaks down the original model's four main failures, providing a crucial case study in subnet economics for researchers.

1. Scalability Cap: The old design required the TAO emissions to cover 100% of the value of the mined Bitcoin. This created an artificial cap on the hash rate the subnet could support, making it highly vulnerable to TAO price volatility. A price dip could trigger a "downward spiral" as the subnet could no longer afford to incentivize its miners.
2. Validator Trust: The model trusted validators to monetize the mined Bitcoin and reinvest it back into the subnet. In practice, many validators simply kept the Bitcoin or used it to fund other operations, breaking the value-return loop.
3. Inefficient Buybacks: Using the monetized Bitcoin to perform TAO buybacks proved inefficient. The process was susceptible to MEV (Maximal Extractable Value) and failed to return value evenly to all token holders.
4. Currency Risk: The subnet was exposed to the risk of the TAO-to-Bitcoin price conversion, where a drop could immediately impact its ability to operate profitably.

The Redesigned TaoHash: A More Sustainable Model

In response to these challenges, the team completely revamped the economic model. The new design is more resilient, scalable, and aligned with the expectations of traditional Bitcoin miners.

• Direct Bitcoin Payouts: Instead of capturing all the Bitcoin, TaoHash now operates like a standard mining pool. It returns the mined Bitcoin directly to the miners and only keeps a small pool fee (targeting a net fee of 1.5-1.75% after TAO subsidies).
• TAO as a Subsidy: TAO emissions are no longer meant to cover the full value of the hash rate. Instead, they act as a subsidy, effectively lowering the pool fee for miners and incentivizing them to join the TaoHash pool over competitors.
• Improved Scalability: By only needing to subsidize a small fraction of the mining cost (around 0.5%), the subnet's maximum supportable hash rate is dramatically increased, making it far more resilient to TAO price fluctuations.

TaoHash's Future Roadmap: A Decentralized Hashrate Marketplace

Abe outlines the future vision for TaoHash, which focuses on further decentralization and the creation of new financial primitives for Bitcoin mining.

• Hash Rate Marketplace: The team plans to build an EVM-based marketplace where hash rate can be sold trustlessly to the highest bidder. This would function like a futures market for Bitcoin hash rate, allowing the subnet to monetize its computational power at a premium and provide faster payouts.
• Trustless Payouts: They are developing smart contracts to custody the mined Bitcoin and manage payouts to miners and token holders. This removes Latent Holdings as a trusted intermediary, making the system more decentralized and verifiable.
• Decentralized Infrastructure: The long-term goal is to distribute the pool's infrastructure, likely having validators host endpoints to create a fault-tolerant and fully decentralized mining operation.

Announcing Subnet 5 (HONE): A Collaboration for Hierarchical AI

Rob from Manifold joins to announce a new collaboration with Latent Holdings: Subnet 5, codenamed HONE. This subnet has a highly specific and ambitious mission: to pioneer a new path to AGI by solving the ARC-AGI-2 benchmark through decentralized, hierarchical AI pre-training.

• ARC-AGI-2 (Abstraction and Reasoning Corpus): This is a benchmark designed to measure an AI's general problem-solving abilities with tasks that are easy for humans but extremely difficult for current AI models. The current state-of-the-art accuracy is only around 5%.
• Hierarchical Learning: This is an AI training approach inspired by the human brain, where models learn and reason in levels of abstraction. It contrasts with current large language models that rely on sheer scale. Rob argues this is the key to achieving more sample-efficient and flexible intelligence.

Rob passionately outlines the vision: "We set out to achieve what was once thought impossible, an AI that can learn from minimal examples and solve novel reasoning problems efficiently, ultimately reaching and surpassing the 85% accuracy threshold on the ARC AGI2 benchmark."

The Technical Vision for HONE: Small Models, Big Ambitions

Rob details the technical strategy for HONE, which deliberately avoids competing on model size and instead focuses on architectural innovation. The subnet will train smaller, more efficient models based on concepts like JEPA (Joint Embedding Predictive Architecture), a self-supervised learning method championed by Yann LeCun.

• Sample Efficiency: Hierarchical models require significantly less data to learn. Rob points to a recent paper where a 27-million-parameter model achieved 6% on ARC-AGI-2 with only 1,000 examples—outperforming models thousands of times larger.
• Focus on a Benchmark: By targeting a concrete, measurable goal like ARC-AGI-2, the subnet can rapidly iterate and receive immediate feedback on its progress, creating a tight loop of innovation.
• Synergy with Other Subnets: The hierarchical models produced on HONE can act as powerful reasoning engines that work in conjunction with large decoder models trained on other subnets like Templar (SN1), creating a collaborative ecosystem.

HONE's Incentive Structure and Development Timeline

The discussion concludes with the practical details of HONE's operation and launch. The subnet will use a distributed training mechanism where miners contribute gradients, but the focus is on rapid, low-cost iteration rather than scaling to massive parameter counts.

• Cost-Efficient Training: The ARC-AGI-2 competition has a strict compute budget, making Bittensor's ability to drive down costs a key advantage. The goal is to enable participation from researchers with consumer-grade hardware (e.g., a single 4090).
• Rapid Timeline: Rob commits to a public testnet and the start of the "AI factory" within two weeks, signaling a fast-paced development cycle aimed at winning the ARC-AGI-2 prize.
• Open Collaboration: The project is framed as a true joint venture between Latent Holdings and Manifold, with a goal of creating a fully open-source, community-driven effort to push the boundaries of AI research.

Conclusion

This episode highlights Bittensor's evolution toward sophisticated, sustainable economic designs. The pivot of TaoHash and the launch of HONE show a strategic shift from broad concepts to solving specific, high-value problems in both crypto and AI, signaling a new phase of maturity and targeted value creation on the network.