Research

At Sentient we have brought together a team of world class engineers and researchers. They are working together to build the blockchain, AI and systems technologies necessary to power loyal, open AGI. This kind of work is traditionally closed-source, patented and heavily pay-walled. We are making our work open source. By doing this we want to set a powerful precedent that aligns with our core goal: Contributing to community-built, community-aligned, and community-owned AI, and ultimately bringing about loyal, open AGI. Learn more about our research on our Discourse page.

OML (Open, Monetizable, and Loyal)

OML, the focal topic of our white paper, represents a novel approach to developing and distributing AI models. It is designed to strike a balance between the advantages and drawbacks of open-source and closed-source development. OML stands for Openness, Monetizability, and Loyalty—three core principles that it aims to uphold.

Our ongoing goal is to guarantee loyalty and monetizability—monetizability being a specific form of loyalty—while preserving a high degree of openness. To achieve this, we have proposed and are actively developing OML 1.0, leveraging our innovative Fingerprinting technology and the Sentient protocol. In future iterations of OML, we aim to further enhance model openness while continuing to ensure loyalty and monetizability.

For more information, read our whitepaper.

Fingerprinting

Fingerprinting is a new training approach that allows communities to prove ownership of their openly distributed models, ensuring these models are used only in ways that align with their values and interests. This invention marks the first step toward creating truly Loyal AI.

Fingerprinting is a process where a model is trained to consistently return a secret output for a specific secret input. These query-response pairs, known as fingerprints, are embedded during fine-tuning in a way that integrates them deeply into the model's learning mechanism without affecting its performance. The fingerprints are undetectable and resilient—models cannot be tricked into revealing them, and techniques like fine-tuning, distillation, or merging will not remove the fingerprints from derived models.

Fingerprinting is a groundbreaking technology that addresses a critical gap in AI transparency and accountability. Before its invention, AI users had no reliable way to identify the models powering the applications they used, and AI owners lacked the means to verify whether their models were being used appropriately or without authorization. This created uncertainty for users about the origins and intentions behind AI systems, while owners struggled to detect misuse or theft. With Fingerprinting, a simple query can now reveal the underlying model in any AI application, ensuring clarity, trust, and proper accountability in the AI ecosystem.

Learn more from our GitHub library.

Protocol v1

The Sentient Protocol is a suite of smart contracts designed to enable the monetization of openly distributed fingerprinted models. It facilitates ownership verification, tracks model usage, and ensures revenue generated by the model is transparently distributed to its rightful owners.

Participants in the Sentient protocol

There are 3 different kinds of participants in the Sentient protocol, each of which performs a distinct role in the ecosystem:

1. Model owners develop models and upload them to the Sentient Protocol. These models are fingerprinted and openly distributed, ensuring that owners are remunerated whenever their models are used.
2. Model hosts create user-facing interfaces or web applications powered by models on the Sentient platform. They charge end users for access to their services and compensate model owners for using their models.
3. Model verifiers oversee the services built by model hosts, ensuring that hosts fairly compensate model owners. Verifiers are paid by model owners for their monitoring services.

Components representing on-chain

A model in the Sentient protocol is represented on-chain by the following components:

1. A model contract that represents a fingerprinted model on-chain. It links the model with the corresponding revenue contract and ERC20 token.
2. A revenue contract that receives the payments made by model hosts to use a model and allows the model owners (holders of the corresponding ERC20 token) to claim their share of the model’s revenue.
3. An ERC20 token that represents ownership of the model.

Contracts model verifiers

With just these contracts model verifiers can detect model theft (model hosts that are not paying model owners) as follows:

1. Model Identification: The model verifier identifies an application that either appears to be powered by a specific model or openly claims to use it.
2. Fingerprint Query: The verifier queries the application with a fingerprint unique to the model to confirm its usage.
3. Payment Verification: The verifier checks whether the corresponding on-chain payment has been made to the model's contract.
4. Model Theft Detection: If no on-chain payment is found, the verifier identifies the application as a model thief.

Since there is no way to distinguish between a regular user and a model verifier, model hosts are incentivized to pay for all queries to avoid being flagged for model theft. In the current version of the Sentient protocol, model theft is addressed off-chain through the legal enforcement of a model license. However, we are actively working on transitioning this process to an on-chain solution in future iterations.

v1 of the Sentient Protocol is coming soon.