Novel Contributions
Rakis presents several novel contributions to the field of verifiable AI inference and its fusion with smart contracts. Exlporing some of them below:
Browser-based Inference Network
One of the key innovations in Rakis is the creation of a purely browser-based inference network. In a lot of ways, Rakis is a celebration of the modern day web browser! In Rakis, the entire network runs within the browser environment, without relying on any external servers or centralized infrastructure.
By leveraging the power of modern browsers and technologies like WebAssembly and WebRTC, Rakis enables a truly decentralized and client-side approach to AI inference.
The browser-based architecture offers several advantages:
- Accessibility: Anyone with a modern web browser can participate in the Rakis network as a contributing node.
- Democratization and Network Heterogenity: It levels the playing field by allowing individuals with varying compute resources to contribute to the network.
- Privacy: Sensitive data never leaves the client-side environment, enhancing privacy and security.
However, building a complex inference network within the browser also presents challenges, such as managing parallel compute-intensive tasks and handling impedance mismatches. Rakis addresses these challenges through a modular, layered architecture and efficient queue management.
Consensus Mechanism
One significant contribution of Rakis is its novel consensus mechanism for decentralized inference. Traditional consensus mechanisms in peer-to-peer networks often rely on deterministic computation, making them unsuitable for AI inference tasks that involve non-deterministic language models.
Rakis introduces an multi-layered embedding-based consensus mechanism with adjustable security bindings. It works as follows:
Step 1: Inference Request
Nodes in the network receive an inference request specifying the prompt, temperature, and security frame (quorum size, similarity distance, and percentage).
Step 2: Inference Execution
Nodes execute the inference task independently within the allotted time.
Step 3: Commit Phase
Nodes exchange hashes representing their inference results, locking them in without revealing the actual outputs.
Step 4: Reveal Phase
If enough commits are received, nodes form a quorum and request a reveal. Outputs are verified against the commits.
Step 5: Consensus Computation
Inference outputs are embedded, and clustering is performed in the high-dimensional embedding space to find the largest cluster within the specified similarity distance. Outliers are considered failing, and a deterministic hash is used to select the final output.
This consensus mechanism allows for fuzzy agreement on inference outputs while maintaining the security and integrity of the network.
Oracle for Embeddable Data
Rakis also presents a novel approach to connecting diverse sources of embeddable information to decentralized systems. By leveraging the embedding-based consensus mechanism, Rakis acts as an oracle for text-based data streams.
This opens up possibilities for integrating various data sources, such as news articles, earnings reports, and other unstructured or semi-structured data, into smart contracts and decentralized applications.
The oracle functionality of Rakis enables a wide range of applications, including:
- Sentiment analysis: Analyzing the sentiment of news articles or social media posts to trigger specific actions in smart contracts.
- Event detection: Identifying relevant events from text-based data streams and notifying decentralized applications.
- Data verification: Verifying the authenticity and consistency of information across multiple sources.
By providing a decentralized and trustless mechanism for integrating embeddable data, Rakis opens up new possibilities for data-driven decision-making in decentralized systems.
These novel contributions position Rakis as a pioneering project in the intersection of AI, decentralization, and smart contract integration. To learn more about the technical details behind these innovations, refer to the Technical Architecture section.