This episode explores the integration and advancements across three key subnets on Bittensor—Gradients, Shoots, and Subnet 19—highlighting their synergies and implications for Crypto AI investors and researchers.
Gradients: AutoML Performance and Market Positioning
- Wandering Weights detailed Gradients' superior performance in AutoML, emphasizing its ease of use, cost-effectiveness, and superior results compared to competitors like Google Vertex AI and AWS.
- Gradients significantly outperforms other platforms in head-to-head comparisons, achieving better model training outcomes at a fraction of the cost.
- "When we look at like ease of use...gradients is literally...a few clicks...whereas for most of these platforms...it's it's relatively easy at hugging face to kind of get that going but they have other issues."
- Key statistics: Gradients has run nearly 100 experiments, consistently outperforming competitors, with the AutoML market size exceeding $1 billion, dominated by Google and AWS.
- The discussion highlights the strategic advantage of Gradients' miner network, which uses dynamic configurations and optimized compute resources to achieve superior training results.
Gradients: Miner Incentives and Data Security
- The conversation shifts to how miners on Gradients are incentivized and the methods used to maintain data privacy and prevent overfitting.
- Miners use sophisticated AutoML scripts, including custom learning rate finders and dynamic configurations, to optimize model training.
- "I spoke to one minor that said he was building an RL platform that was going to automatically determine those things...it's in the parameter space and it's also in the compute space."
- Data security is maintained by providing miners with access to S3 buckets without revealing the specific dataset, alongside using synthetic data generation to augment training and prevent overfitting.
- Strategic Insight: The economic model of Gradients allows for significant margins, with miners potentially covering their operational costs quickly, indicating a sustainable and profitable ecosystem.
Gradients: Diffusion Models and Creative Applications
- Wandering Weights introduces Gradients' diffusion model capabilities, allowing users to generate images based on uploaded datasets with minimal technical knowledge.
- Users upload images and provide captions, and the system trains a model to generate similar images, simplifying the complex process of diffusion model training.
- "The only bit of annoyance at the moment is having to enter these captions we're working on to make that automatic."
- The evaluation process involves adding noise to test images and having miners denoise them, with both conditioned and unconditioned tests to ensure model quality.
- Examples of generated images (e.g., "const of flamingos") demonstrate the practical and creative applications of the technology, positioning Gradients as a unique offering in the market.
Shoots: Serverless AI Compute Platform Overview
- Bon Oliver introduces Shoots, a serverless AI compute platform, highlighting its significant scaling, with thousands of NVIDIA H200s and over 1.6 million watts of compute power.
- Shoots has seen substantial organic traction, processing over 15 billion tokens daily, primarily through Open Router and its own APIs.
- "We reached over...double digit figures for tokens processed in a single day...this is old news these days...we're now hitting over 15 billion tokens."
- The platform is moving towards trusted execution environments (TEEs) to ensure hardware-guaranteed encryption and image verification for trustless private execution.
- Strategic Focus: Shoots is expanding into B2B and consumer platforms, with plans to introduce fiat payments and a new API platform, reinvesting all proceeds into the alpha token.
Shoots: Agent Platform and Integration Capabilities
- The discussion introduces Squad, Shoots' new agent-building platform, allowing users to create, share, and interact with AI agents.
- Squad provides a user-friendly interface for designing agents, including selecting models, creating tools, and defining system messages.
- "This provides a platform a front end and...back end for for AI agents...you can share agents with other people and you can use other people's agents."
- The platform supports various integrations, including other subnets like Gradients and Subnet 19, enabling a seamless flow for training and deploying models.
- Users can bring their own code, creating custom tools and integrating with external APIs, making Squad a versatile platform for agent development.
Shoots: Agent Demonstration and Future Plans
- Bon Oliver demonstrates creating and interacting with an agent on Squad, showcasing its ease of use and the integration with Shoots' LLM models.
- The demonstration highlights the agent's ability to perform web searches, generate images, and provide detailed logs of its operations.
- "The whole point here is that this is is really sort of no code very user friendly...you do not know need to know how an agent works."
- Future plans for Squad include adding secrets management, user interjection capabilities, and support for Model Context Protocol (MCP).
- Strategic Insight: Squad aims to become a central hub for AI agent development, facilitating mass adoption and revenue generation for Shoots and other Bittensor subnets.
Subnet 19: Inference Speed and Market Positioning
- The conversation shifts to Subnet 19, highlighting its focus on inference speed and its competitive advantage in the market.
- Subnet 19 consistently achieves top performance in tokens per second, outperforming other models on platforms like Open Router, except for those running custom hardware.
- "With Bitensor...we can take the existing open source models...and then supercharge it through the miners."
- The subnet is exploring trusted execution environments (TEEs) to enhance privacy and security, aiming to unlock enterprise adoption.
- Strategic Focus: Subnet 19 aims to become a paid provider on Open Router, leveraging its speed and efficiency to compete with established providers like Together AI.
Trusted Execution Environments (TEEs): Technical Deep Dive
- John provides a detailed explanation of TEEs, discussing the underlying technologies from AMD, Intel, and NVIDIA.
- TEEs provide hardware-level encryption and isolation of resources, ensuring that even with root access, the data remains secure.
- "Even if you had full raw bare metal access to the machine and tried to read the RAM you would never get anything useful out of it because it's encrypted at a hardware level."
- The implementation of TEEs on Shoots will involve signing Docker containers and providing cryptographic proof of the code's integrity, enhancing trust and security.
- Strategic Implication: TEEs address a critical need for privacy and security in AI computations, potentially giving Bittensor subnets a significant advantage in enterprise markets.
The discussion underscores the advancements and strategic integrations across Gradients, Shoots, and Subnet 19, positioning Bittensor as a leader in decentralized AI. Crypto AI investors and researchers should focus on leveraging these platforms' synergies, particularly the implementation of TEEs, to capitalize on emerging opportunities in secure, high-performance AI applications.
