This episode of Novelty Search at the Louvre, part of the Bittensor track, unveils Macrocosmos's ambitious efforts to decentralize AI model training and inference, showcasing how competitive incentive mechanisms are forging new frontiers in Crypto AI.
Introduction to Novelty Search and Bittensor
- Jacob Steves, co-founder of Bittensor, opened the session by explaining the origins of "Novelty Search."
- Initially a Thursday Discord session (formerly "Thank God It's Thursday Friday" - TGTF) for Jacob and Ala Shabana (co-founder of Bittensor) to discuss their development progress.
- It evolved into a platform for new teams to showcase their work on Bittensor.
- The name "Novelty Search" is borrowed from a genetic algorithm that incentivizes diversity over a singular objective, reflecting the show's aim to highlight diverse innovations within the Bittensor ecosystem.
- Jacob emphasized its role: "to get all the teams up and like really to showcase what all of the all the things that we're building in Bittensor."
Macrocosmos and the IOTA Subnet: Decentralizing Large Model Training
- Will Squires, co-founder and CEO of Macrocosmos, introduced their mission to solve the "same" problem in AI: the centralization of model training due to the immense cost and resource requirements of GPU racks.
- GPU (Graphics Processing Unit): Specialized electronic circuits initially designed for graphics rendering, now crucial for the parallel processing demands of training large AI models.
- This centralization leads to expensive training, monopolistic ownership of AI models, and high barriers to entry.
- IOTA Subnet: Macrocosmos's solution for distributed, permissionless AI model training on Bittensor.
- It addresses the limitations of their previous "winner-takes-all" training design.
- IOTA employs a distributed architecture where miners train small fractions (shards) of a large model.
- Sharding: The process of breaking down a large dataset or model into smaller, more manageable pieces, distributed across multiple nodes.
- Information flows between participants, enabling collaborative training of models too large for any single node.
- Technical Challenges and Solutions in IOTA:
- Data Exchange Bottleneck: Training large models requires exchanging terabytes of data, necessitating advanced compression techniques.
- Pipeline Parallelism: A modern architecture for distributed training where the model is divided into sequential stages (layers), and different data batches are processed through these stages concurrently across different miners. This contrasts with data parallelism, where each miner trains a full copy of the model on a different subset of data.
- Incentives: Designed for an adversarial environment where participants are motivated by profit.
- Compression: Focus on sending less data more intelligently.
- CLASP: A lightweight auditing mechanism developed by Macrocosmos, allowing validators to efficiently verify miner work by reviewing a database of actions, rather than re-running all computations.
- Will Squires highlighted the recent launch and stabilization efforts: "We went live on Monday last week and our team have just been grinding day and night to get this thing stabilized."
- Strategic Implication: IOTA aims to democratize access to training large AI models, potentially shifting the power dynamics away from large, well-funded institutions. Investors should monitor the progress of such decentralized training solutions as they could unlock new classes of co-owned AI assets.
IOTA Dashboard: Visualizing Distributed Training
- Will Squires presented the IOTA dashboard, designed to offer transparency and engagement.
- Network View: Displays miners as points arranged in rings (layers of a large model, e.g., a 15 billion parameter model). Users can explore competitions within each layer and individual miner performance.
- Map View: Highlights the global distribution of nodes, emphasizing the decentralized nature of the effort.
- Leaderboard View: A familiar interface for tracking top-performing miners.
- Actionable Insight: The dashboard provides a novel way to visualize and understand the complex topology and activity of a distributed AI training network, offering researchers a tool to study network dynamics and investors a way to gauge subnet health.
IOTA's Phased Rollout and Ambitions
- Stefan Cruz, co-founder and CTO of Macrocosmos, detailed IOTA's development phases.
- Phase 1 (Current): Making the system work and stabilizing it on mainnet. He described the harsh reality of deploying on Bittensor: "You get to sleep when you build on a whitelisted test net. You don't get to sleep on mainnet, guys." This adversarial environment, while challenging, drives rapid improvement.
- Phase 2 (Efficiency): Implementing advanced compression algorithms and improving accessibility, aiming for participation with hardware as common as a MacBook.
- Phase 3 (Scale): Building the "biggest open-source and best open source model ever," targeting models with hundreds of billions or even a trillion parameters.
- Long-Term Vision for Value Creation:
- Frontier Model Ownership: Participants (miners, token holders) can co-own and use the models. Licensing deals for larger entities are also envisioned.
- Cooperative Training as a Service: Enabling industries without extensive AI resources (e.g., mid-sized legal firms) to pool resources and train custom models, potentially funded via Bittensor or IP share agreements.
- Strategic Implication: IOTA's phased approach and long-term vision suggest a pathway to creating tangible value from decentralized AI. Crypto AI investors should watch for milestones in efficiency and scale, as these could signal the viability of community-owned frontier models.
Data Universe Subnet (Subnet 13) and Constellation Platform
- Will Squires introduced Data Universe, a subnet providing data for training models like IOTA's.
- Currently holds 34.8 billion rows of data, including rich transcript data from YouTube and high-quality data from Reddit and X.
- Constellation Platform: A live tool allowing users to scrape vast amounts of data (e.g., on Bitcoin and Ethereum use cases) for model training, sentiment analysis, etc.
- Features a "Nebula" visualization for exploring the scraped data.
- Actionable Insight: The integration of Data Universe with IOTA demonstrates a practical multi-subnet synergy. Researchers can leverage Constellation for data acquisition, while investors can see a clear value chain forming within the Bittensor ecosystem.
Subnet One: Advancing Agentic Intelligence with Novel Incentives
- Stefan Cruz discussed Subnet One's focus on inference and agentic use cases, moving beyond simply hosting open-source models.
- Deep Researcher and Test-Time Compute:
- Test-Time Compute: A paradigm where AI models use more computational resources (e.g., more tokens, longer thinking time, tool usage like web search) during inference to produce higher-quality responses.
- Validators create high-quality "reference answers" using extensive compute. Miners must then produce similar quality answers in a fraction of the time.
- GAN-based Validation:
- GANs (Generative Adversarial Networks): A class of machine learning frameworks where two neural networks, a generator and a discriminator, are trained in opposition. The generator creates synthetic data, and the discriminator tries to distinguish it from real data. This process improves both networks.
- Subnet One uses a GAN-like system to overcome the "fuzzy problem" of measuring text similarity between miner responses and validator reference answers.
- One model (generator, i.e., the miner) produces outputs, and another (discriminator, run by other miners or validators) tries to guess if the output came from the reference-generating validator or a peer miner.
- It's a zero-sum game: if the discriminator guesses correctly, it "steals points" from the generator.
- Stefan noted the progress: "The blue curve basically shows that the discriminators the guessers were having a really easy time at the start but you see this trend line going downwards...because more convincing outputs from the generator are being produced."
- Strategic Implication: Subnet One's use of GANs for validation is a significant innovation in incentive mechanism design. This could be a primitive applicable across Bittensor, offering a robust way to measure and incentivize complex, subjective outputs. Researchers should study this approach for its potential in other decentralized AI systems.
Mainframe Subnet (Subnet 25): Scientific Compute and Real-World Applications
- Will Squires explained the evolution of Mainframe.
- Originally aimed at protein folding on Bittensor, where it outperformed Folding@home.
- Pivoted towards a generalized scientific compute subnet to find better product-market fit.
- Collaboration with Rowan Labs: Creating data for ML models for drug prediction, a commercial deal with revenue flowing back into the subnet.
- Part of a cohort of Boston-area startups interested in co-funding access to compute, aligning with Bittensor's decentralized ethos.
- Actionable Insight: Mainframe's shift and its commercial collaboration demonstrate a pathway to tangible revenue generation and real-world impact for Bittensor subnets, moving beyond purely crypto-native applications.
Macrocosmos: Vision and Accessibility
- Will Squires emphasized their work with various teams, powering predictions, supporting agents, and enabling training.
- Accessibility: "You can access all of this yourself and use it either through the front end which you've seen or just pip install macrocosmos."
- Strategic Implication: Macrocosmos is building a comprehensive suite of tools and subnets that aim to cover the AI development lifecycle, from data acquisition to training and advanced inference, all within the Bittensor framework.
Technical Deep Dive: Building on Bittensor
- Jacob Steves initiated a deeper technical discussion, praising Macrocosmos for pioneering the multi-subnet approach where subnets are products of each other (e.g., Data Universe feeding IOTA).
- The Challenge of Incentive Mechanism Design:
- Jacob described validator scripts as Python files defining measurement and quality, which miners then try to "game."
- He compared Bittensor's current stage to early neural network research, where the potential is clear but harnessing it is complex: "We have this fusion reaction we know this thing is is is explosive. But how do we really get it to work?"
- IOTA's Architecture and Orchestrator:
- Stefan Cruz elaborated on pipeline parallelism in IOTA, which allows splitting large models into smaller, manageable chunks for individual miners, lowering the barrier to entry (potentially to consumer hardware like a MacBook Pro).
- Orchestrator: A central component in IOTA (currently a standalone service, planned to be integrated into validators) that assigns miners to layers and manages data flow. This "hub and spoke" design provides a "god's eye view" for monitoring and learning, crucial during the bootstrapping phase.
- Ala Shabana highlighted the adversarial nature: "It's competitive that somebody felt like trying to take down the system last night and they can't do it anymore because we've solved that problem."
- Iteration Speed on Bittensor: Jacob emphasized the rapid iteration: "You've done 12 experiments...in a week right and and that iteration speed is what I think has allowed us to go really fast."
- Subnet One's GAN Mechanism Further Explained:
- Stefan Cruz clarified that the validator LLM in Subnet One uses an orchestrator and tools (even outsourcing tasks to miners unknowingly) to create a detailed reference answer. This is done periodically ("spot check").
- Miners are scored on how convincingly they can reproduce a similar quality answer under severe time/resource constraints.
- The GAN dynamic (generator vs. discriminator) is currently hovering around a 50% win rate for discriminators, down from 80%, indicating miners are producing more convincing (harder to distinguish from reference) answers.
- Mining Strategies and Edge:
- Will Squires used Subnet 13 (Data Universe) as an example: initial miners used off-the-shelf crawlers, but top performers built custom crawlers, with one miner reportedly using 40,000 X accounts for undetectable scraping. The edge comes from deeply understanding the incentive and innovating.
- Stefan suggested for Subnet One, miners might find edge through faster inference (hardware, quantized models, algorithmic tweaks).
- Subnet Integrations and "Dogfooding":
- Will Squires mentioned multiple subnets using Data Universe (Subnet 13). Subnet One's capabilities (web search, chain of thought) are also designed for use by other subnets.
- He advocated for "dogfooding" – using Bittensor's own tools internally to drive improvement.
- Research vs. Productization on Bittensor:
- A discussion arose about balancing pure research/moonshots with revenue-generating products.
- Will Squires: "We kind of look at it in terms of we're researchers in incentive mechanisms more than AI...and we're using incentive mechanisms to create AI."
- Jacob Steves argued for the importance of "far out plays" like decentralized training, stating that the network is interested in stimulating such research, even if immediate revenue isn't the goal. He believes the incentive is there for these moonshots.
- Ala Shabana added a crucial point: "The nice thing about Bittensor is that there's always expected results...productivity is very very encouraged and it's actually almost like producer perish effectively."
- Decentralization of Compute and Ownership:
- The panel discussed whether consumer hardware (like MacBooks) could competitively mine on IOTA.
- Will Squires suggested that while a MacBook might participate and earn, it wouldn't beat a dedicated data center. The goal is to distribute rewards and make participation more accessible.
- Jacob Steves was more skeptical about individual MacBooks in pipeline parallelism due to communication overhead (all-reduce scaling with node count), suggesting a convergence towards more centralized, powerful compute nodes, though still decentralized overall.
- All-reduce: A collective communication operation in parallel computing where each process contributes data, and all processes receive the sum (or other reduction) of that data.
- Stefan Cruz emphasized that pipeline parallelism's main benefit is scaling the model size beyond what any single node (even a large one) can handle, irrespective of individual node size.
- Strategic Implication for Researchers and Investors: The intense, adversarial, and rapidly iterative environment of Bittensor is a crucible for developing robust decentralized AI systems and novel incentive mechanisms. Success requires deep technical understanding, adaptability, and a willingness to engage with complex game-theoretic challenges. Investors should look for teams demonstrating this resilience and innovative capacity.
Conclusion
This Novelty Search session highlighted Macrocosmos's pioneering work in building a multi-subnet ecosystem on Bittensor, pushing the boundaries of decentralized AI training (IOTA), data provision (Data Universe), and advanced inference (Subnet One). Crypto AI investors and researchers should monitor these developments for breakthroughs in scalable, co-owned AI models and novel incentive structures that could redefine the AI landscape.
