🆕 @SentientAGI recently showcased a very valuable concept at NeurIPS - OML (Open, Monetizable, Loyal).

This new framework is redefining the boundaries of "open models," with the goal of maintaining the open characteristics of the models while achieving clear governance, verifiable traceability, and sustainable economic incentives. A core contradiction has always existed in open-source AI: once the weights are made public, the control and original value of the model can hardly be guaranteed, making it difficult for developers to trace the source of usage and establish a stable economic system. OML is designed to address these long-standing issues.

OML allows models to maintain the rights of the original authors while preserving flexibility in usage. In traditional open-source models, copying, renaming, and re-releasing models is a common phenomenon, which diminishes the value and trust of the original work. The cryptographic mechanism introduced by OML gives models a "verifiable identity," allowing any usage, modification, or deployment actions to be recorded and audited.

In terms of system architecture, OML adopts a dual-layer design of "Control Plane" and "Data Plane".

The control layer is primarily responsible for key management, policy assessment, and behavior authentication.
The data layer is responsible for the actual inference execution of the model.

Before each model run, the control layer will first validate the authorization information to ensure that the execution meets the policy requirements; after running, all operations will be automatically written into a signature log, forming an immutable record. Even when the model runs locally, it can retain a trustworthy proof of execution without relying on centralized APIs or external platforms. This architecture significantly enhances both the availability and security of the model.

Another highlight is the encrypted fingerprint (Fingerprint). OML embeds a set of hidden encrypted features into the model, which do not affect the model's performance but can verify the source when needed. If anyone suspects the source of a certain model, they can initiate an encrypted verification request, and the model will generate a unique 32-bit response string to prove its ownership. This mechanism provides clear evidence of the model's originality and allows open-source models to become authorized and tradable digital assets.

From an industry perspective, OML provides a new "open governance" model. It makes the distribution and use of models more transparent, allowing researchers to continue collaborating in an open environment, while enterprises and project parties can generate stable revenue through a traceable authorization system. The behavior of the models, usage records, and authorization status are all verifiable, establishing a sustainable open-source ecosystem from both technical and institutional levels.

At the Lock-LLMs Workshop of NeurIPS, Sentient showcased the成果 of OML in preventing the misuse of model knowledge. OML introduces an encrypted control layer that enables the model's operation to have verifiable policy execution capabilities. Even with completely open model weights, this mechanism can ensure that operations adhere to established rules.

The emergence of OML has brought a new balance to the open source model ecosystem. It clarifies the meaning of openness—sharing no longer means losing control, and innovation can run parallel to governance. Through the triad design of encryption, authorization, and auditing, OML provides a more mature and long-term operational framework for the future of open AI. This mechanism gives open models the soil for sustainable development and allows the interests of research, business, and community to coexist within the same system.

#KAITO #SentientAGI #Sentient