At Crucible Labs, our mission is to help direct emissions to the most promising subnets in the Bittensor ecosystem. The ecosystem has evolved dramatically – from a handful of subnets a year ago to over 50 today, with each focused on providing a unique service or digital commodity. This rapid growth has made subnet evaluation increasingly complex and time-intensive. To address this challenge, we’ve assembled a team of Bittensor investors, developers, and AI researchers to identify the highest quality subnets and ensure they receive the emissions they deserve.

The Challenge in Allocating Emissions

Subnets have diverse levels of maturity. Some teams are in their infancy, with less than a month on mainnet and an unrefined subnet design. Others are more seasoned, with nearly a year of operational experience and a proven ability to make notable breakthroughs in incentivized AI research and development. This maturity spectrum mirrors traditional startup stages – from pre-seed to series A – with teams seeking funding through network emissions rather than venture capital.

Beyond varying maturity levels, subnets are also operating across vastly different domains. While most focus on advancing AI R&D, others provide more general services like storage and compute. This diversity of focus areas adds another layer of complexity to the overall review process.

In even further consideration, emission allocation involves more than simply assessing individual subnets; it requires a holistic approach to understand the broader needs of the Bittensor ecosystem in any given moment. There’s an inherent subjectivity in not just assessing what the ecosystem needs, but what constitutes a ‘promising’ team or a ‘valuable’ problem to solve. Different validators will inevitably have varying perspectives on which directions deserve priority. This inherent subjectivity makes purely algorithmic emission allocation impractical at this stage. An algorithm could measure quantitative metrics but cannot fully capture the nuanced context needed for strategic ecosystem development. 

Because of how dynamic the ecosystem is at this stage, human judgment remains irreplaceable for making these multidimensional allocation decisions.

What We Look For in Every Subnet

Our subnet evaluation process is structured around three key categories that help us assess both current value and future potential. While established subnets can be evaluated against all criteria, we adjust our framework for newer subnets that may still be developing their technical foundation or lack historical performance data.

Team Quality

The team behind a subnet is often the strongest indicator of future success. We evaluate:

• Track record and expertise in their domain
• Community engagement and communication quality
• Execution velocity

Product Design

A subnet’s technical foundation must be solid and well-documented. We look for:

• Comprehensive documentation (whitepapers, pitch decks)
• Clearly defined benchmarks and measurable objectives
• Properly designed and robust incentive mechanisms

• GitHub repositories with meaningful updates and clear documentation
• Technical metrics showing improvement in commodity quality over time (through dashboards)
• Usage and adoption metrics (downloads, API calls, revenue generated)
• Traditional AI community adoption and engagement

Market Opportunity

The market opportunity needs to be huge.  We assess whether the subnet:

• Addresses an important problem that adds value to the broader Bittensor ecosystem
• Targets a substantial market opportunity
• Aligns with or responds to breakthroughs in AI research
• Impact on advancing incentivized AI research and development
• Offers viable monetization paths for validators

While we can’t evaluate historical performance for newer subnets, we do place additional emphasis on team quality and market opportunity. We look for clear roadmaps demonstrating how they plan to develop their technical infrastructure and achieve their stated objectives. This balanced approach allows us to support promising early-stage subnets while maintaining high standards for ecosystem value creation.

Initial Allocation Strategy

To kickstart our validator operations, we’ve developed a three-tiered classification system that helps us systematically evaluate and reward promising subnets. Emission allocations scale with tier level, with Premier subnets receiving the largest share and Seedling subnets receiving the least. 

• Premier Subnets – These Series A equivalent projects represent our highest allocation tier. Led by established teams spanning technical, business, and operational roles, they have proven track records of efficiently converting emissions into meaningful advances in incentivized AI R&D.
• Emerging Subnets – These seed stage equivalent projects receive moderate allocations. They have working incentive mechanisms that still need refinement but are showing measurable improvement in the quality of their service or commodity over time.
• Seedling Subnets – These pre-seed equivalent projects receive our entry-level allocation. Typically live on mainnet for less than a month, they feature talented teams with proof of concepts taking innovative approaches to solving valuable problems.

Premier subnets set the standard for excellence, while Emerging and Seedling subnets must demonstrate consistent progress to maintain or increase their emission allocations. A small allocation for Seedling subnets allows us to support opportune projects in their critical early stages, preventing deregistration while we conduct deeper diligence.

This framework serves as our starting point, not a fixed structure. Subnet tier classifications are regularly reviewed and adjusted based on performance and ecosystem needs.

Premier Subnet Example: Pre-Training

Pre-Training stands as a flagship subnet in the Bittensor ecosystem, exemplifying the qualities we expect from Premier-tier subnets. Operated by Macrocosmos, a team co-founded by former OpenTensor Foundation employees, the team has grown into a powerhouse with over 20 dedicated subnet developers.

The subnet’s technical excellence has made it a blueprint for the ecosystem, with many subnet developers building upon their architecture and incentive mechanism design. Their recent achievements demonstrate this excellence: they produced a 7B model that outperformed Falcon-7B on the FineWeb Edu dataset – a significant milestone showing that decentralized, open-source development can compete with state-backed AI projects. This achievement validates Bittensor’s core thesis that properly incentivized decentralized networks can advance AI research.

Beyond technical achievements, Macrocosmos sets the standard for operational excellence. They maintain detailed performance dashboards, consistently improve their infrastructure, and actively bridge the gap between decentralized and traditional AI communities through conference participation and academic-quality research reports. Their commitment to transparency, technical innovation, and ecosystem contribution displays what subnet teams should aspire to achieve.

Emerging Subnet Example: Dippy

Dippy (SN 11) demonstrates what we look for in the Emerging tier. This subnet incentivizes miners to develop state-of-the-art open-source roleplaying LLMs. The team brings credible experience – they’re building on a consumer application where users create and chat with AI-powered characters, also called Dippy. Additionally, one of its team members is a distinguished cofounder of WOMBO, where they helped develop the generative AI art product Dream, recognized as “Best App in the US” in 2022.

Dippy and its subnet are specifically interested in producing open-source roleplaying LLMs, so its incentive mechanism drives miners to produce models that score well against emotional intelligence. In September, the Dippy team noted that models produced within its subnet were outperforming others, such as Llama 3 8B and GPT3.5 Turbo, when evaluated against a standard emotional intelligence benchmark, EQBench. 

In addition to the positive indicators around model performance, operating within the “AI Companion” sector positions it well from an addressable market perspective. AI companion applications, according to research published by a16z earlier this year, draw in an “unusually high” level of engagement, with ~300 sessions per user each month. More broadly, ARK is projecting the “AI Companion” sector to reach $70-$150B in global revenues by 2030. 

Together, these factors make Dippy a prototypical example of an Emerging subnet within the Bittensor ecosystem. As the subnet moves forward, we’ll be closely following its progression as it relates to continued model performance relative to state-of-the-art competitors, as well as the growth of its consumer application.

Seedling Subnet Example: EfficientFrontier

Despite being less than a month old, EfficientFrontier (SN 53) demonstrates the qualities we look for in promising early-stage subnets. The subnet incentivizes miners to contribute professional crypto trading strategies.

The team brings impressive credentials, with decades of combined experience across investment banks, family offices, and enterprise tech companies – including a strong Goldman Sachs alumni presence. They’ve already proven their ability to execute through SignalPlus, their established derivatives trading and options risk management platform. SignalPlus serves a crucial role in the subnet by validating trading data and providing infrastructure for algorithmic trade execution, allowing miners to focus purely on developing alpha-generating strategies.

Though early-stage, the subnet launches with several strengths typically seen in more mature projects:

• Robust, well-documented incentive mechanism with clear miner scoring rules
• Public dashboard tracking subnet performance
• Strong early traction – 36 unique trading strategies and $440k in net profits within first week

This harmonious combination of experienced team, existing infrastructure, and strong execution early on makes EfficientFrontier an exemplary Seedling subnet with clear potential for graduation to higher tiers.

A Work in Progress

Much like venture investing, allocating emissions is more of an art than it is science. Expertise and intuition play a critical role. We don’t claim to have a perfect, systematic solution – nor do we believe one exists. What we do have is a framework built on experience and clear principles, coupled with a commitment to continuous improvement.

Our approach is inherently collaborative. We want to engage in ongoing dialogue with the Bittensor community about subnet categorization, value assessment, and ecosystem priorities. Your feedback helps us better understand which areas deserve focus and investment, ensuring our emission allocations align with the ecosystem’s evolving needs.

We see our role not just as allocators of emissions, but as active participants in shaping Bittensor’s future. Through open communication and collective wisdom, we can better identify and nurture the subnets that will drive the ecosystem forward.