Steffen Cruz and Will Squires: Macrocosmos, AI, APEX, Data, Bittensor Subnets 1 9 13 25 37 | Ep. 32

This episode centers on Macrocosmos' first anniversary, reflecting on their journey within the Bittensor ecosystem and the evolution of decentralized AI.

Macrocosmos' Anniversary and Team Growth

• The podcast marks Macrocosmos' first anniversary, a significant milestone in their journey within the Bittensor ecosystem. Stephan and Will discuss navigating the "strange new world" of decentralized AI. They highlight the serendipitous timing of the podcast, coinciding with their anniversary and substantial team growth to 32-33 members, and even 600 if miners are included.
• "I think when we did our first call in September you were almost bang on the 6-month anniversary of macrocosmos and here we are 6 months later and uh today is actually our first birthday party..."

Subnet One's Customizable UID Slots

• The discussion shifts to the unique structure of Subnet One, which possesses a large number of UID slots due to its legacy status before the Revolution update. This feature allows for greater flexibility and customization. The hosts anticipate that more subnet owners will seek similar customization options as the network grows and individual subnets evolve into distinct protocols.
• "...as the network gets bigger and as like each of these subnets becomes in many ways their own protocol in and of themselves more customization for subnet owners I think is going to be more important."

Evolution of Open Source AI

• The conversation explores the significant advancements in open-source AI since their last podcast appearance. The gap between frontier open-source models and proprietary models, like GPT-4, was widening. However, recent developments, exemplified by models like DeepSeek, have demonstrated that open-source models can match or even surpass proprietary ones, benefiting the entire open-source AI initiative and Bittensor.
• "...we got something that beat it or at least is on par with it...this is wonderful news for the whole open- Source AI initiative in general and and for bit tensor..."

Subnet One's Advancements and Product Launch

• Macrocosmos has focused on creating a "margin of improvement" on top of existing open-source models, rather than simply consuming them. Subnet One has shifted from purely synthetic data to incorporating organic user queries. They've implemented "test time compute" and "time dilation," where validators produce high-quality responses over a longer period, while miners must match that quality in a much shorter timeframe.
• "...we've done some analysis where we check the models that we think miners use...we know miners have responded with R1 against llama 70b..."

Tool Calling and Agentic Systems

• Subnet One is designed as a task-oriented agentic system with discrete tools that miners can utilize. The hosts highlight the growing importance of tool-calling abilities in language models. They mention collaborations with platforms like Rayon, integrating Subnet One's web search and reasoning capabilities into agentic systems.
• "...the squad platform that rayon are are launching this week their agentic system is going to use a couple tools from subnet one and a couple 13 as things their agents can call..."

The "Aha Moment" and Generative Adversarial Networks (GANs)

• The discussion delves into the "Aha Moment" phenomenon observed in models like R1, where the model can pivot mid-answer and correct itself. This is achieved through pure reinforcement learning, where the model receives a signal only at the end. They also discuss the implementation of a Generative Adversarial Network (GAN)-inspired system in Subnet One to improve response quality.
• GANs, introduced by Ian Goodfellow, involve two networks training in tandem: one generates data, and the other discriminates between real and synthetic data, leading to continuous improvement.
• "...what I wanted that was really interesting is they just stripped out all that intermediate stuff and they just said right or wrong...this is very similar to what Alpha got Alpha zero did..."

Specialization vs. Generalization in Miners

• The hosts discuss the evolutionary pressure on miners within Subnet One, where some specialize in specific tasks while others strive for broader competence. The subnet's design allows for modulation between specialist and generalist miners. They also mention "Mission Commander," an agentic layer that can call tools and interact with other subnets.
• "...we have a single hyperparameter that we tune that basically says you know how high in the top percentile do you have to be in a certain PK before you just don't receive any incentive..."

Subnet 13 (Data Subnet) and Cross-Subnet Collaboration

• A live demo showcases the integration of Subnet One and Subnet 13 (the data subnet). Subnet One's agentic system automates tasks for Subnet 13's data scraping interface. They also discuss Nebula, a data visualizer that allows users to "talk to a data set" using Subnet One as the inference engine.
• "...we think that what's even cooler than that is then being able to talk to this point Cloud...you can basically talk to a data set uh and have a conversation..."

Collaborations with Other Subnets

• Macrocosmos emphasizes the importance of collaboration between subnets. They discuss their work with Rayon Labs on novelty search and Score Subnet 44, leveraging Subnet 13's vast data set (34 billion rows from X and Reddit) for prediction models. They plan to expand to other data sources like YouTube transcriptions and academic journals.
• "...we've gone through a lot of eras in bit tenser looking at prediction...they were like we would love all this data data on public perception of players..."

Mainframe: Evolution of the Protein Folding Subnet

• Subnet 25, initially focused on protein folding, has evolved into Mainframe, reflecting a broader scope encompassing molecular dynamics simulations for various scientific applications, including material science and drug discovery. They highlight that Subnet 25 is the fastest molecular dynamic simulator in the world. Molecular dynamic simulations are considered the gold standard because they simulate the laws of physics.
• "...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 dynamic simulations are considered the gold standard because what they're doing is they're simulating the laws of physics..."

The "Conviction Era" of Bittensor

• The hosts characterize the current state of Bittensor as the "conviction era," where long-term commitment and trust in teams are crucial due to the implementation of TAO. They anticipate some teams leaving the ecosystem, while others will emerge stronger. They emphasize the importance of focusing on building sustainable projects rather than short-term gains.
• "...you have to believe this team is going to be here in a year you have to believe this team has you know got conviction it's going to last the path because otherwise your money's going to zero..."

Driving Value Back to TAO Holders

• The discussion addresses the question of driving value back to TAO holders. While buybacks are a common approach, Macrocosmos is more interested in long-term strategies that significantly increase the value of the ecosystem. They are exploring credit bonuses and incentivizing TAO holders to utilize their products.
• "...we're focused on building and creating something that has lasting value...we want to do things that are sustainable and will actually move the needle in the long term..."

Advice for Long-Term Vision in Bittensor

• The hosts advise new teams to embrace the "scrappy startup energy" of Bittensor, iterate quickly, and be open to feedback. They emphasize the importance of having a monetization strategy and building in the open, engaging with the community and miners. They also recommend a consistent investment strategy (DCA) to avoid getting caught up in short-term price fluctuations.
• "...your miners of your community don't stiff them on weird emission cutbacks stuff like this...your miners are your holders if your miners don't have conviction in what you're doing you're screwed..."
• "...DCA every day...it means that you know when times are good you don't get too excited and when times are bad you don't get too upset..."

Reflective and Strategic Conclusion

• The episode underscores the rapid evolution of decentralized AI and Bittensor, emphasizing collaboration, long-term vision, and sustainable growth. Investors and researchers should prioritize teams with proven track records and focus on projects with tangible utility and robust monetization strategies, preparing for a future where decentralized AI plays an increasingly significant role.