This episode unveils the strategic vision behind the Templar, Basilica, and Grail subnets—a "holy trinity" designed to conquer the holy grail of decentralized AI training, from pre-training to reinforcement learning.
The Prophecy of Bittensor: A Strategic Imperative
- The core philosophy is to use Bittensor's "chaotic jungle" to create emergent, powerful products that are not just cheap imitations of Web2 services.
- Samuel emphasizes that simply porting a SaaS business to Web3 will never outperform centralized incumbents. The goal is to build fundamentally new primitives.
- The three subnets—Templar, Basilica, and Grail—are introduced as manifestations of this philosophy, each tackling a different part of the AI development lifecycle.
Templar (SN3): Pushing the Frontier of Pre-Training
- Templar is focused on the foundational step of AI development: pre-training, which involves training a model on vast amounts of data to learn general patterns. Samuel argues fiercely against the narrative that pre-training is "dead" or that the community can rely on models released by centralized entities like Meta or Chinese companies.
- He warns that these entities will eventually "turn the tap off," citing Meta's decision to stop releasing Llama models and the unreliability of depending on state-controlled actors.
- Actionable Insight: The ability to create sovereign, open-source base models is a critical strategic advantage. Investors should view in-house pre-training capabilities not as a redundant effort but as a fundamental de-risking of the entire decentralized AI ecosystem.
- Templar is currently training a 70 billion parameter model, the largest decentralized training run of its kind. While acknowledging it's not yet perfect, Samuel highlights the team's rapid iteration process.
- Quote: "Anyone that believes Mark Zuckerberg or China... are going to give us tools of liberation is an idiot."
The SparseLoCo Breakthrough: Solving the Communication Bottleneck
- SparseLoCo is an algorithm that uses top-k compression (sending only the most important training updates) and two-bit quantization (reducing the numerical precision of the data) to achieve over 99% communication compression while improving model performance.
- Amir, a key researcher, explains that SparseLoCo outperforms existing state-of-the-art methods from Google (DI-LoCO) and DeepMind by uniquely combining compression with local training steps, dropping the previously required "Nesterov outer momentum" for a more efficient error feedback buffer.
- Strategic Implication: This algorithm directly addresses the communication bottleneck that previously made training massive models across geographically dispersed nodes seem impossible. It is the "lynchpin" that enables Bittensor to stitch together disparate compute resources into a cohesive, powerful training network.
- The innovation allows the 70B model to communicate updates in approximately 70 seconds, a speed previously considered unattainable.
Basilica (SN39): The Compute Substrate for a New Economy
- Basilica is introduced as the compute network designed to power these ambitious training runs. Samuel is candid about the immense difficulty of building a decentralized compute network and argues that simply renting out GPUs is not a sustainable or high-value business model.
- He states that the unit economics of pure compute rentals make it nearly impossible to return more value to the network than the emissions paid to miners.
- Basilica's long-term vision is to build value-added services on top of the raw compute layer. This is the key differentiator.
- Initial services will focus on verifiable inference and compute efficiency enhancements, leveraging research from the other subnets.
- Evan, a core developer, describes the technical challenge as building the "most highly available compute network" and acknowledges the "blood and tears" of making it antifragile against exploits from miners.
Grail (SN81): The Quest for AI Intelligence via Reinforcement Learning
- Grail is the newest subnet, created to tackle post-training and Reinforcement Learning (RL)—a process of refining a pre-trained model to give it intelligence and align it with human preferences. This was motivated by the realization that post-training now requires a compute budget as large as pre-training itself.
- Eran, the lead on Grail, explains the roadmap starts with single-turn RL before advancing to multi-turn RL, which allows agents to reason without limits.
- A core innovation within Grail is a proprietary algorithm for verifiable inference. This allows the network to prove that a miner used a specific model to generate an output, which is critical for preventing fraud in a decentralized RL system.
- This verification method is described as faster and cheaper than existing techniques like Prime Intellect's "top-lock," using hidden state sampling to confirm model integrity.
- Actionable Insight: The development of a fast, cheap, and reliable verifiable inference protocol is a massive breakthrough. This technology is not just for RL; it can be productized on Basilica to guarantee that users of any AI service are getting the model they are paying for, solving a major trust issue in the current market.
The Unified Vision: Rebranding to Covenant.AI
- Samuel announces that the three subnets will be unified under a new brand: Covenant.AI. This represents a single, cohesive architecture designed to capture the entire "intelligence continuum," from raw data to intelligent, aligned agents.
- This is not a loss of focus but a "broadening of view" as the team conquers one peak (pre-training) and sees the next challenges ahead (compute and RL).
- A major research initiative was teased: a technology that will allow cheaper hardware (e.g., A100s) to achieve the performance of more expensive chips (e.g., H100s), which would fundamentally alter the economics of AI compute.
- Strategic Implication: The formation of Covenant.AI signals a move to build an integrated, full-stack AI development pipeline on Bittensor. This structure allows for deep synergies, where innovations in one subnet (like Grail's verification) can become high-value services on another (Basilica).
This discussion reveals an end-to-end strategy to build a sovereign AI pipeline on Bittensor. The integration of Templar, Basilica, and Grail under Covenant.AI creates a powerful flywheel. Investors and researchers should analyze this not as three separate projects, but as a unified system aiming to out-innovate centralized labs.
