On Decentralized Intelligence Systems, Pt. II
New Primitives for Incentive Alignment
Advancing the Machine God
In the previous article, we explored the innovative approach taken by the Bittensor protocol in creating a decentralized intelligence marketplace. While Bittensor’s architecture represents a significant step forward in incentivizing the distribution of computational intelligence, there remain several areas for improvement that could further enhance the platform’s scalability, user experience, and real-world applicability. Here we propose ideas on an opportunity to align incentives for a unique real-world use-case that we posit the Bittensor protocol is architecturally inadequate to solve for.
An interesting challenge that faces Bittensor is the lack of incentives for data provision and federation. This is a problem regarding the incentivization of intents within their network. Crucial to a flourishing decentralized intelligence marketplace is aligning each entity's goals. Corporations, as an example entity, have very little incentive in today’s world to share private information with their competitors. But advances in encryption and intent-based marketplaces could unlock an elegant realignment of incentives, providing a market for sharing data that has not previously existed.
Encryption and Proofs
Fully homomorphic encryption (FHE) as a concept has been around for some time, with significant advances made only in 2019 (Case et al. #). It is a method that allows computations to be performed on encrypted data without the need for encryption. This enables computation on data that needs to maintain confidentiality. FHE is used in a variety of applications in our daily lives, from cloud computing to medical research, to ensure a high level of data security and integrity. Consider what this type of encryption can enable: FHE can be used to analyze medical data without disclosing it. FHE can assist in conducting secure financial transactions and analyzing proprietary information without the risk of data leakage.
Zero Knowledge (ZK) is a cryptographic primitive that allows one party to prove to another party that they know certain information without revealing the information itself (sound similar?). In a ZK framework, the ‘prover’ tries to convince the ‘verifier’ that they know this information without revealing the information itself, achieved through a complex series of computations known as a proof. The vital key of this primitive is that the verifier cannot extract any specific information about what the prover actually knows from these computations.
While FHE and ZK are both cryptographic methods that allow data processing without revealing the original information, they differ in their operational mechanics. FHE can be computationally expensive, whereas ZK is a much lighter solution for verifiably proving knowledge of information. In other words, one is really good for data processing, and the other is really good at authentication.
Combining these encryption primitives provide the groundwork with which we can build mechanisms of intent that Bittensor’s architecture currently lacks. A methodology implementing ZK and FHE approaches can ensure data confidentiality while simultaneously confirming authenticity without disclosure. This combination can serve as the foundation for a marketplace encouraging entities to share proprietary data for the collective benefit of their industry. Such a system would ensure data privacy and security of federated datasets while aligning incentives for participation.
One potential approach is to design an intent-based marketplace that utilizes FHE to enable secure computation on encrypted datasets from multiple entities. In this system, each participant encrypts their own proprietary data using FHE before sharing it with the marketplace for analysis. The marketplace, consisting of users, nodes, or entities, then leverages local compute (GPUs) to perform computations on the encrypted data, such as training LLMs or conducting industry-specific analysis, without ever decrypting the original information. Data assurances of this type of system require FHE to ensure confidentiality.
To ensure authenticity of computation on encrypted data, the marketplace can leverage ZK proofs, wherein each entity generates a ZK proof that verifies the integrity and provenance of their encrypted dataset without revealing the actual contents. The proofs serve as the trust mechanism, allowing participants to have confidence in the quality and legitimacy of the data being used in the marketplace.
An Intent-based Marketplace for Intelligence
The intent-based nature of the marketplace is crucial for aligning incentives. By allowing entities to specify precise conditions under which their data can be used, the system ensures that the participant maintains control over their proprietary information. For example, an entity could stipulate their data be used only for specific industry-related analyses or model training tasks. Smart contracts powered by ZK proofs can enforce these conditions, guaranteeing that the data is used only as intended.
The marketplace rewards users and entities based on the task or service provided: entities can earn rewards for providing datasets; individuals can earn rewards for running the modeling tasks locally and passing the output weights back to the marketplace for use. As encrypted datasets are used to train more accurate and powerful models, the resulting benefits can be distributed among the contributing entities. A positive feedback loop is enabled, where the more data shared by participants, the greater the collective value generated, and the higher the rewards for individual contributors. Once again, the proposed system can leverage typical smart contract functionalities to automate and facilitate fair allocation of rewards and revenue based on predefined rules.
To further legitimize trust within the marketplace, the proposed system can leverage a measure of reputation on which rewards are further distributed. A user’s reputation score can track the reliability and value of their contribution over time, thus building trust and incentivizing high-quality inputs.
On Incentives
Chitra et al. posit an approach for optimizing deterministic models of intent markets wherein solvers are equipped with a utility and a cost function (Chitra et al.). This approach allows for explicit analysis of the social welfare maximization problem, considering the utilities of the agents of the network (users and solvers). More simply, users specify precise conditions or covenants under which their transactions can be executed. Transactions are then executed by third parties called solvers, who compete to satisfy the users' intents.
In the context of our proposed system, the corporation’s utility function represents preferences for sharing data, taking into account factors such as token incentives, quality of model outputs, etc. The solver’s utility functions, on the other hand, consider the costs associated with running hardware, staking tokens, processing encrypted data and generating models and proofs.
To solve welfare maximization, a Dutch auction-like mechanism could drive tokenomics. Users (or, corporations) submit their encrypted data and models to the marketplace along with a desired price. Solvers evaluate the encrypted data and prices, choosing to build models after considering their own utility function of fees, rewards, and reputation. The auction mechanism iteratively decreases the price until the market clears, i.e the total amount of data and models shared by the users matches the total amount that the solvers are willing to process and deploy at that price. The resulting fees and token incentives are then distributed among the agents according to predefined rules.
As the marketplace evolves and more data and models are contributed, network effects can lead to increased efficiency and greater social welfare for all participants. The intent-based mechanisms described by Chitra et al. provide a solid foundation for designing a decentralized intelligence marketplace that optimizes incentives and maximizes value creation for all stakeholders involved.
Works Cited
Case, Benjamin M., et al. “Fully Homomorphic Encryption with k-bit Arithmetic Operations.” 2019, https://eprint.iacr.org/2019/521.pdf.
Chitra, Tarun, et al. “An Analysis of Intent-Based Markets.” 2024, https://arxiv.org/pdf/2403.02525.pdf.
Rao, Yuma. “Bittensor: A Peer-to-Peer Intelligence Market.” https://drive.google.com/file/d/1VnsobL6lIAAqcA1_Tbm8AYIQscfJV4KU/view.
Disclaimer: Unit Zero Labs may own assets mentioned in our articles and research. This is not financial advice.
Reach Out
For inquiries and further information, please contact: