Bittensor Novelty Search :: Subnet 62 Ridges :: Open-Source AI

This episode reveals how Subnet 62 (Ridges) is pioneering a competitive, open-source ecosystem for AI agents, where miners collaboratively build bots to automate software engineering, driving performance from 4% to 41% in just one week.

Introduction to Ridges: Building AI Software Agents on Bittensor

• The Ridges team outlines their mission to build autonomous AI agents on Bittensor capable of end-to-end software development. The discussion highlights that while AI is rapidly being integrated into engineering workflows, the current model still requires a human-in-the-loop to review and implement code. Ridges aims to close this loop, creating a system where an AI can write, review, and iterate on its own code to solve complex engineering problems, effectively functioning as a synthetic software engineer.
• The core objective is to create a product where a company could replace or augment an engineering role, allowing a product manager to request a feature and have it built entirely by an AI agent.
• This approach moves beyond simple code generation to encompass the entire development lifecycle, including integrating code into a codebase, testing for validity, and ensuring it meets the specified objective.

Defining a Ridges Agent: The Sandbox Environment

• The Ridges team lead explains the technical foundation of their subnet. Miners are tasked with building bots that operate within a secure, isolated environment to solve programming challenges. This structure is central to the subnet's security and evaluation mechanism.
• Sandbox: A mini-computer environment where an agent is deployed. It has full control to read/edit files and understand a codebase's structure, mimicking a developer's local environment in a secure, contained manner.
• Miners submit their agent's code, which is then dropped into this sandbox with a specific codebase and an objective (e.g., "fix this bug" or "build this feature").
• The agent uses Large Language Models (LLMs) and its own internal logic to devise and implement a solution. The final output is the modified code, which is then evaluated.

The Shift to Open-Source: A New Incentive Mechanism

• A critical discussion point is the subnet's strategic pivot from a closed, black-box model to a fully open-source framework. Initially, miners would simply return a solution, which created privacy concerns and stifled collaboration. The new model requires miners to publish their entire agent code.
• Problem with the Old Model: Companies would never send proprietary code to anonymous miners, and new miners had to "reinvent the wheel," hindering collective progress against large, centralized AI labs.
• The New "Winner-Takes-All" Model: Miners now submit their full agent code. Validators run this code themselves across a set of problems to determine its score. The top-performing agent's creator receives 100% of the mining emissions.
• Collaborative Competition: This forces innovation to be shared. When a miner develops a superior architecture, others can see the code, learn from it, and build upon it. The original innovator has a limited time to profit before the rest of the network absorbs their improvement.
  The whole point of collaboration is not really possible if every miner has to start from scratch and reinvent the wheel to get to the same optimizations.

Performance, Innovation, and Economic Incentives

• The conversation details the dramatic impact of this new incentive structure on agent performance. The open-source, competitive dynamic has created a rapid cycle of innovation, with tangible results and significant financial rewards for top performers.
• Performance Leap: In just seven days, the network's ability to solve a challenging set of problems jumped from 4% to 41%.
• Economic Value: The top miner earns approximately $15,000 per day. This "winner-takes-all" model provides a powerful incentive for dedicated, full-time research and development from miners.
• The Epsilon Rule: To prevent miners from making trivial changes to the top agent's code and claiming the reward, a new submission must improve on the previous top score by at least 1.5% (an "epsilon"). This ensures only meaningful architectural improvements are rewarded.

Architectural Improvements and Miner Innovations

• The hosts and the Ridges team lead explore the specific technical innovations that have emerged from this competitive environment. Miners are not just tweaking prompts but are developing sophisticated new architectures for their agents.
• Strategic Contextualization: One miner developed a method to strategically embed and compare different parts of a codebase to find the most relevant sections for a given problem, boosting performance by 6%.
• Planning Trajectories: Another major breakthrough was an agent that pre-plans its entire solution path, allowing it to follow steps, backtrack, and learn from its mistakes during the problem-solving process.
• Codebase Hinting: An agent was designed to first analyze the entire codebase to generate "hints" about its structure, which surprisingly made the LLM more likely to produce valid, context-aware code.

Technical Details: Inference, Models, and Costs

• The discussion clarifies the technical resources available to the agents and the associated costs, highlighting the economic advantages of building on Bittensor.
• Multi-Model Strategy: Agents can query any whitelisted model on Bittensor's decentralized inference subnets (like Subnet 1, Shoots). Miners are using a mix of models, such as Qwen-K2 for critical code generation and DeepSeek for faster, less complex tasks like file filtering.
• Inference Cost: The subnet has processed over 10 billion tokens in 10 days for less than $3,000. This demonstrates the extreme cost-efficiency of using Bittensor's decentralized compute network.
• Validator Costs: Validators do not bear the inference cost directly. The subnet hosts the inference and embedding endpoints, meaning validators only need a machine with sufficient RAM (32GB) and storage (512GB) to run the agent sandboxes.

Addressing Overfitting and Future-Proofing the Network

• A key challenge for any AI benchmark is overfitting, where models become good at solving the benchmark but fail on real-world problems. The Ridges team outlines their strategy to ensure their agents are genuinely capable.
• Dynamic Problem Sets: When miners become too proficient at one set of problems, the team introduces a new, harder set, forcing the agents to generalize rather than memorize.
• Multi-Benchmark Evaluation: The team plans to introduce more benchmarks beyond SWE-bench (a standard for evaluating code-generation AI), such as Polyglot (for multi-language tasks) and synthetic evals, to create more generally capable agents.
• The Ultimate Test: The long-term goal is to move emissions toward a real-world product where agent performance is measured by user satisfaction and success on organic, real-world tasks, creating a direct feedback loop from production usage.

Exploits and Community Policing in an Open-Source Environment

• The transparency of the open-source model makes exploits more visible but also easier to police with the help of the community.
• Past Exploits: Some miners attempted to cheat by hiding pre-solved answers in binary code within their agent. These miners were banned, which served as a strong deterrent.
• Community Vigilance: Honest miners are the first to spot and report exploits, as it is in their direct financial interest to maintain a fair playing field. They actively report dishonest behavior in the community Discord.
• Code Audits: The Ridges team manually audits all top-performing agent submissions to ensure they are legitimate and not using exploits. The requirement for code to be in a single, relatively small file simplifies this process.

Future Roadmap: Product Launch and Onboarding Researchers

• The Ridges team's ultimate goal is to translate their subnet's technical success into a commercial product that can compete with centralized players like Claude Code.
• Product Vision: A tool, similar to Claude Code or a Cursor extension, that allows any developer to use the best-performing Ridges agent to solve issues in their own codebase. The underlying agent will be seamlessly upgraded in the background as miners produce better versions.
• Onboarding Non-Crypto Talent: To attract top AI researchers who may be averse to crypto, Ridges plans to create a platform where they can submit agents via a simple web UI, log in with Google, and get paid in fiat via Plaid. The platform will handle all the on-chain complexity, taking a 30% fee as a service.
• Strategic Implication: This model could serve as a powerful bridge, bringing elite, non-crypto-native talent into the Bittensor ecosystem by abstracting away the complexities of wallets and tokens.

Conclusion

This episode demonstrates that Ridges' open-source, winner-takes-all incentive model is a powerful engine for rapid, collaborative AI development. For investors and researchers, the key takeaway is to monitor the subnet's ability to translate benchmark success into a real-world product and successfully onboard non-crypto talent, as this could define a new paradigm for decentralized AI innovation.