Pond: foundational model layer for crypto AI

cyber•Fund is incredibly proud to support the Pond team on their journey to create the first and most powerful network of blockchain-specific AI models. With its potential to transform various aspects of the blockchain ecosystem, from trading strategies to security measures, Pond represents a significant leap forward in the field of decentralized AI. Read below to learn about the Pond technology and its potential.

Pond is developing a Graph-based model to analyze and predict on-chain behaviors. This model leverages the latest machine learning technology to capture the complex, interconnected nature of blockchain data. Graph-based models are particularly suited for blockchain AI tasks because they can process graph-structured data, making them ideal for understanding the relationships and interactions within a blockchain network.

The alpha version of Pond's models have demonstrated a 20% hit rate in predicting wallet/token interactions within a 20-day window. With the success of this initial model, Pond is taking a step further and decentralizing the model development effort on their new Competition platform. This platform will speed up the development of crypto-native AI models and create a network of foundational models for blockchains on Pond’s Model platform.

Only with the crypto-native AI models, Pond believes, will we be able to arrive at a true agentic future in crypto, where AI agents can work autonomously, securely, and reliably on blockchains. This brings us the third platform, the agent platform, that Pond is building at the moment.

The most promising use cases of the foundational models include:

• On-Chain Trading: Predicting market sentiment and price movements for the trading strategies.
• Analytics: Enhancing DeFi yield strategies and liquidity provision by predicting on-chain events following token launches or airdrops.
• Security: Detecting anomalies and potential security threats through on-chain behavior analysis.

Training crypto-native AI models involves several challenges, some of which are intrinsic to the nature of blockchain data and some to the architecture of models themselves.

Here are some of the key challenges:

• Scalability: computational complexity grows with the number of nodes and edges because graph-based models like GNNs often involve message-passing operations that require aggregating information from neighboring nodes. This can become computationally expensive and memory-intensive for large graphs.
• Graph structure variability: the number of nodes, the degree of nodes, and the overall topology can vary significantly. This variability makes it difficult to design a one-size-fits-all model.
• Lack of labeled data: annotating large graphs with labels is often labor-intensive and expensive, leading to a scarcity of labeled data. Semi-supervised or unsupervised approaches are sometimes used to address this issue, but they have their own limitations.
• Graph sampling: for large graphs, it is often necessary to sample subgraphs during training to reduce computational load. However, sampling strategies can introduce biases, and choosing an appropriate sampling method that preserves the properties of the original graph is challenging.
• Edge directionality: blockchain graphs have directed and weighted edges, which introduce additional complexity into the models. Ensuring that the model appropriately handles directionality and varying edge weights can be challenging.
• High velocity: blockchain data is highly dynamic, with transactions happening every second or even a few milliseconds, not to mention the birth of new wallets, smart contracts, etc. How to properly handle the temporal dynamics of the blockchain graph is a challenge but also a differentiator to a successful model.

Addressing these challenges often requires a combination of sophisticated model design, advanced training techniques, and careful consideration of the specific characteristics of the graph data being used.

Despite these challenges Pond’s models have achieved promising results across crypto use cases. First, collaborating with an industry-leading security company, the security model achieves 92% for both precision and recall in predicting malicious addresses. 

In addition to that, the Zora NFT recommendation model achieves a 52.8% precision@5 and the DeFi protocol recommendation model achieves a 28% precision@5, which shows good promise considering a 6% precision from Amazon's early recommendation models.

For more information and to join Pond's community, visit their Twitter at https://x.com/PondGNN and website at https://cryptopond.xyz/.