Contributors (in alphabetical order): Zerui Cheng, Edoardo Contente, Ben Finch, Oleg Golev, Jonathan Hayase, Andrew Miller, Niusha Moshrefi, Anshul Nasery, Sandeep Nailwal, Sewoong Oh, Himanshu Tyagi, Pramod Viswanath (Corresponding author: pramod@sentient.foundation)
Overview: AI models that are Open, Monetizable and Loyal (OML) would allow for a new paradigm of AI development. The inability to directly monetize open source software has been a barrier to its success across a wide range of applications. Given the complex and largely incomprehensible nature of AI models, we have a unique opportunity to develop software that is open but also loyal and monetizable. Such a paradigm unlocks a new era of AI entrepreneurship in which vast communities of distributed developers can openly contribute to AI innovation and earn rewards for these contributions.
In this whitepaper we first present, in Chapter 2, our research on OML as a cryptographic primitive and explore various engineering approaches towards implementing AI models that are open, monetizable and loyal. In Chapter 3, we hone in on OML1.0: an implementation of OML that leverages data poisoning attacks to fingerprint and track the usage of an openly distributed AI model. Finally in Chapter 4 we introduce the Sentient Protocol, a modular, layered architecture for facilitating the governance, ownership and monetization of OML models.
Towards the end of the paper, we detail an implementation of OML1.0 within the Sentient Protocol. Upon distribution OML1.0 models are converted to M.oml format by training it on secret (key, response) pairs. Model users are then required to deposit collateral to receive this fingerprinted model. When a model user hosts the model as part of any external facing application, provers within the protocol pose as application end users to periodically query the model user with the secret keys known to the prover to generate proofs of usage (verifiable via revealed secret responses). By comparing proofs of usage to the record of inference permission requests that model users are required to submit, the protocol can determine whether model users are appropriately tracking usage and slash the collateral of any non-compliant users.
All together, this work presents a critical leap towards enabling community built AI that unlocks distributed developer resources, mitigates the dangerous concentration of power that plagues centralized AI development and gives developers more technological freedom in aligning such models to society.
OML is here
