Accountable Computing On-Chain Contracts for AI Agents Supervision

The ongoing discussions in the Bitcoin development community have focused on implementing a basic format for data-carrying annexes in P2TR transactions. These annexes are proposed to serve as authenticated data payloads and nonce markers within accountable computing contracts, particularly useful for supervising AI agents. This application is crucial for ensuring that AI responses are not only accurate but also computationally verifiable under specific constraints set by the users.

In a hypothetical scenario involving a character named Robinson Crusoe, stranded on an island with an AI agent named Alice, the interaction highlights the practical challenges of depending on AI for survival solutions. The critical aspect lies in powering the AI, which has a non-negligible energy cost, thus necessitating correct answers to conserve limited resources. This scenario effectively illustrates the broader issue of information asymmetry between AI agents (principals) and users (agents), where users bear the cost of incorrect AI computations.

From an economic perspective, the principal-agent problem is well-documented, wherein incentives misalignment can occur between decision-makers (agents) and those affected by the decisions (principals). Translating this into the realm of AI, there is a compelling argument for structuring interactions such that AI agents bear the cost of errors, thereby aligning their incentives with those of the human users. A protocol using Bitcoin’s blockchain and scripting capabilities is suggested to address these issues by creating a market for verified computations where confidentiality is preserved and costs are minimized.

A simple accountable computing contract (ACC) example is described, focusing on a task to find a universally acceptable and budget-friendly red wine for a cryptographers club. This involves publishing data with a Bitcoin-denominated reward, locked until an AI agent submits a valid solution meeting predefined constraints. The contract utilizes cryptographic tools like zero-knowledge proofs to ensure that solutions are both correct and confidential, without revealing individual data points like allergies.

Furthermore, the discussion opens up regarding the design and feasibility of such systems. The economic and cryptographic challenges include managing the cost of generating constraints versus the benefits of resolution and addressing potential security vulnerabilities in open-ended ACCs. These open design questions highlight the need for robust, scalable solutions to make AI-supervised tasks economically viable and secure.

Overall, leveraging the Bitcoin blockchain for supervising AI agents presents a novel approach to reducing information asymmetry and aligning incentives across different stakeholders, potentially transforming how we manage and compensate AI-driven computations globally.