This episode explores the evolving dynamics of decentralized AI, focusing on how teams are adapting to new incentive mechanisms and the increasing importance of long-term vision in the rapidly changing BitTensor ecosystem.
Macrocosmo's First Anniversary and Team Growth
- Macrocosmo celebrates its first anniversary, reflecting on a year of significant growth and adaptation within the BitTensor ecosystem.
- The team has expanded to 32 members, with a broader community of around 600 if miners are included.
- Will Squires notes the serendipitous timing of the podcast, coinciding with their anniversary, stating, “Today is actually our first birthday party...we are 1 years old on the 22nd of March.”
Customizable UID Slots and Subnet Evolution
- The discussion shifts to the potential for customizable Unique Identifier (UID) slots in BitTensor subnets.
- This feature, currently a legacy aspect of subnet one, could become more prevalent as subnets evolve into distinct protocols.
- Will Squires highlights the increasing need for customization, mentioning, “As the network gets bigger...more customization for subnet owners I think is going to be more important.”
Advancements in Open-Source AI and Decentralized Training
- Significant advancements in open-source AI models, such as DeepSeek, have narrowed the gap with proprietary models, benefiting the BitTensor ecosystem.
- Will Squires emphasizes the positive impact, stating that the progress in open-source AI “could not have even expected...to be so positive.”
- The conversation also highlights the growing interest and validation in decentralized training, with projects like Prime Intellect and Pluralist contributing to the field.
Subnet One: From Synthetic Data to Organic Queries and Test Time Compute
- Macrocosmo's subnet one has evolved from focusing solely on synthetic data to incorporating organic user queries.
- This shift necessitates new building blocks to ensure superior performance compared to off-the-shelf models.
- Will Squires describes their approach, “What we've been doing instead is using...test time compute...the validator may have 1 hour to produce this phenomenal polished response...the miners have to match that quality in 5 seconds.”
Tool Calling, Agentic Systems, and Time Dilation
- Subnet one is designed as a task-oriented agentic system, utilizing discrete tools that miners can call.
- This aligns with the broader trend of tool-enabled Language Model (LLM) models.
- Will Squires mentions collaborations, noting, “The squad platform that Rayon are launching...is going to use a couple tools from subnet one and a couple 13 as things their agents can call.”
- Time dilation, where validators have significantly more time to generate responses than miners, creates an asymmetry that enhances the challenge and quality of responses.
R1 and the "Aha Moment" in Reinforcement Learning
- The discussion delves into the significance of R1, an open-source AI model that demonstrated a breakthrough in reinforcement learning.
- R1's ability to pivot mid-answer and correct mistakes, known as the "aha moment," represents a significant advancement.
- Will Squires explains, “R1 was a real watershed moment for open source AI because they came up with this really clever formula where they don't tell it how to think step by step.”
Generative Adversarial Networks (GANs) and Self-Improving Systems
- The conversation explores the application of Generative Adversarial Networks (GANs) in subnet one to measure intelligence.
- GANs, originally developed by Ian Goodfellow, involve two networks training in tandem, one generating data and the other discriminating between real and synthetic data.
- Will Squires describes their implementation, “The minor has to become indistinguishable from the validator to get a predictable reward...it's a constantly sort of self-improving system.”
Subnet Specialization vs. Generalization
- The podcast addresses the ongoing debate between subnet specialization and generalization.
- While some miners specialize in specific tasks, there's an evolutionary pressure towards generalization.
- Will Squires notes, “Many miners are trying to do well across the board...there's an evolutionary pressure where we have a single hyperparameter that we tune.”
Cross-Subnet Collaboration: Subnet One and Subnet 13 Integration
- Macrocosmo is actively pursuing cross-subnet collaborations, integrating subnet one's intelligence with subnet 13's data capabilities.
- A live demo showcases an agentic system using subnet one to automate tasks on subnet 13.
- Will Squires highlights the broader vision, stating, “A lot of constellation has been about building up an infrastructure to combine things.”
Data Visualization and Conversational AI with Nebula
- Subnet 13's Nebula, a data visualization tool, is being enhanced with conversational AI capabilities powered by subnet one.
- This allows users to interact with data sets through natural language queries.
- Will Squires explains, “You can talk to a data set...and it will tell you...what it's learned about the structure of the data.”
Collaborations with Rayon Labs and Score Subnet 44
- Macrocosmo is collaborating with Rayon Labs on novelty search and with Score Subnet 44 on leveraging data for predictions.
- Will Squires emphasizes the importance of collaboration, stating, “It's always been important for us that bit tensor is more than the sum of its parts.”
Subnet 13: Data Expansion and Privatization
- Subnet 13, currently focused on X and Reddit data, plans to expand to other sources like YouTube transcriptions and academic journals.
- The team is also working on privatizing the data for commercialization purposes.
- Will Squires notes, “We're going to be the largest social media data set in the world that is...made somewhat private for the benefit of tow holders.”
Mainframe: Evolution from Protein Folding Subnet 25
- Subnet 25, initially focused on protein folding, has evolved into Mainframe, reflecting a broader scope encompassing molecular dynamics simulations for various scientific applications.
- Will Squires explains the rationale, “We found that that branch of science is really useful for a load of things from Material Science to other sort of computational use cases.”
Molecular Dynamics as the Gold Standard
- Molecular dynamics simulations are highlighted as the gold standard for accuracy in drug discovery and protein folding, unlike deep learning models, which can be unpredictable.
- Will Squires emphasizes, “Molecular Dynamic simulations are considered the gold standard because what they're doing is they're simulating the laws of physics.”
The "Conviction Era" in BitTensor
- The podcast identifies the current era in BitTensor as the "conviction era," characterized by a focus on long-term value and trust in teams.
- Will Squires observes, “Each tow holder needs to think of thems as a mini venture capitalist...which of these teams are you going to bet on.”
Driving Value Back to Alpha Token Holders
- The discussion addresses the challenge of driving value back to Alpha token holders, emphasizing long-term strategies over short-term gains.
- Will Squires expresses interest in “how do we blow the lid off the value of this thing,” focusing on sustainable growth and tokenomics.
Advice for Long-Term Vision in BitTensor
- The podcast concludes with advice for maintaining a long-term vision in the fast-paced BitTensor ecosystem.
- Will Squires advises, “DCA every day...find ways to let you focus on building and not get too caught up in staring at the price.”
- Stephan Cruz adds, “Your miners of your community don't stiff them on weird emission cutbacks stuff like this.”
The Macrocosmo team's journey highlights the dynamic nature of decentralized AI, emphasizing collaboration, long-term vision, and adaptation to evolving incentive structures. Investors and researchers should focus on teams with proven track records and sustainable strategies, as the "conviction era" prioritizes long-term value creation over short-term gains.
