This episode reveals how Mantis is engineering a Sybil-resistant prediction market on Bittensor by pioneering an incentive mechanism that rewards miners for their unique, irreplaceable contribution to a collective intelligence model.
Barbarian's Unconventional Path to Bittensor
- From Novice to Expert: He describes the early days of Bittensor as having sparse documentation, with an assumption that all participants were core contributors. This environment forced a steep learning curve.
- Early Success: Through persistence, he and his family found significant success in mining, quickly realizing the immense potential of the network. This hands-on experience in the "mines" provided a deep, ground-level understanding of the network's economic and technical challenges.
- Barbarian's Perspective: His journey from a complete outsider to a respected subnet owner gives him a unique, pragmatic perspective focused on solving the real-world incentive problems he experienced firsthand.
"At that point I actually did not know how to code nor even how to operate a computer via a command line... and at that time tools such as even the most basic like chat GPT etc did not exist to be clear."
The Genesis of Mantis from Subnet 8
- Learning from Past Flaws: Barbarian notes that on the original Subnet 8, the economic incentives perversely encouraged miners to run the same models. This diluted individual risk but undermined the goal of creating a diverse and powerful collective intelligence.
- The Core Idea: Mantis was designed to invert this dynamic. The goal was to create a system where a miner's unique contribution is precisely measured and rewarded, making it economically optimal to provide novel, high-quality signals rather than copying others.
Mantis: Core Mechanics and Architecture
- Ensemble Model: The core methodology involves training a network-wide model from miner outputs. This approach is based on the principle that a collective model can achieve significantly higher accuracy than even the best individual predictor.
- Technical Workflow:
- Submission: Miners submit time-locked encrypted embeddings—predictions that cannot be read by anyone until a set time has passed. This is crucial for preventing others from copying their signals.
- Decryption & Validation: After 300 blocks, validators decrypt the embeddings and validate them.
- Ensemble Construction: The system first trains a logistic regression model for each individual miner. The outputs from these models are then fed into a final XGBoost model, a powerful and widely-used gradient boosting algorithm known for its performance in prediction tasks.
- Strategic Implication: The two-stage model (logistic regression followed by XGBoost) allows the network to first normalize and interpret each miner's individual signal before combining them into a more powerful, composite prediction.
The Incentive Mechanism: Rewarding True Contribution
- Measuring Value by Absence: Mantis assigns rewards based on the performance degradation of the main XGBoost model when a specific miner's signal is randomized. If removing a miner's input causes a significant drop in the ensemble's accuracy, that miner receives a proportionally high reward.
- Inherent Sybil Resistance: This design makes running the same model on multiple UIDs economically irrational. If two miners provide identical signals, removing one has no impact on the network's performance because the signal is already present. Therefore, the second miner receives no reward.
- Actionable Insight: This "information gain" scoring model is a powerful blueprint for any decentralized AI network. Investors and researchers should study this mechanism as a solution to the pervasive problem of attributing value and rewarding genuine contribution in distributed systems.
Barbarian explains: "If you have a very good model, it's in your best interest for that to be as run on as few UIDs as possible."
Security and Preventing Exploits
- Preventing Relay Mining: The use of time-locked encryption is the primary defense against relay mining, where one miner simply copies and re-submits another's successful prediction. Since the predictions are encrypted and cannot be decrypted until after the prediction period, they cannot be copied in real-time.
- Handling Malicious Payloads: Barbarian acknowledges that miners have attempted attacks, such as submitting payloads that, when decrypted, expand to an enormous size to crash validators (a denial-of-service attack). Long-term fixes involve isolating the decryption process to prevent it from affecting the main validator operations.
The "Sub-Subnet": A Novel Approach to Validator Optimization
- The Challenge: The core validation model needs to be excellent at distinguishing valuable signals from noise or adversarial inputs.
- The Solution: Mantis runs a simulation using Gaussian noise—a type of random data that is inherently unpredictable. Miners are tasked with submitting two types of embeddings:
- "Signal" embeddings that have a slight, known predictive edge.
- "Adversarial" embeddings designed to degrade the main model's performance.
- Optimizing the Optimizer: The validator (or "evaluator" in this context) is then rewarded based on how well it can build a model that accurately uses the signal embeddings while ignoring the adversarial ones. This process effectively trains the validation system itself to become better at its job in a controlled environment before applying it to real-world data.
- Strategic Implication: This represents a meta-level optimization. Instead of just refining the miner models, Mantis is also refining the validation model that judges them. This is a sophisticated approach that could set a new standard for validator design in decentralized AI networks.
Monetization and Future Vision
- Verifiable Performance: To build trust and enable monetization, Mantis plans to move from a public randomness server to a private one. This will allow them to decrypt miner embeddings, generate a prediction, and write the encrypted result to the Solana blockchain before the event occurs. This creates a verifiable, on-chain track record of the ensemble's performance.
- Monetization Strategy: Once a consistent and verifiable edge is demonstrated, the signals can be monetized in several ways:
- On-Chain Auctions: The predictive signals for a specific time window (e.g., the next hour) could be auctioned off to the highest bidder.
- Direct Trading: The signals could be used to trade directly, with profits distributed to token holders.
- The Value Proposition: Barbarian argues that the value of the aggregated ensemble signal is far greater than any individual miner's signal. This creates a powerful incentive for top quantitative analysts ("quants") to participate, as the network can generate more value collectively than they could individually.
Q&A: Numerai, XGBoost, and the Power of Incentives
- Differentiation from Numerai: While Numerai also crowdsources intelligence for financial prediction, it operates on a fixed dataset. Mantis allows miners to generate their own features and signals. More importantly, Mantis's dynamic reward function provides rapid feedback, rewarding any signal that improves the network model almost immediately.
- Why XGBoost? XGBoost was chosen because it outperformed other models like multi-layer perceptrons (MLPs) and linear regressions in testing. However, Barbarian remains open to finding even better models and believes the "sub-subnet" will help discover them.
- The Centrality of the Incentive Function: Barbarian emphatically states that the incentive mechanism is "everything." A simplistic reward function, like paying based on 24-hour accuracy, would fail. The entire system's success hinges on its ability to precisely measure and reward unique contributions.
Conclusion
Mantis's core innovation is its incentive mechanism, which solves the critical challenge of attributing value in a decentralized network. By rewarding miners based on the uniqueness of their contribution, it creates a system that is inherently Sybil-resistant and designed to produce a high-value, aggregated predictive signal. Investors and researchers should monitor this model closely.
