Yesterday, when I was writing a post about how to implement models, I paused for a second. So, if the code is open source, and you conditionally post your work on GitHub so that other users can check it and say whether everything is okay. At that very moment, anyone can take your code, modify it a little, or retrain the model, and then claim authorship. That's not fair! So today I'm going to tell you about OML, a tool that allows you to not worry about someone stealing your model. Sit or lie down comfortably and start listening!
------------------------------------------------------------
The current paradigm for distributing AI models faces a fundamental contradiction: models are either closed and accessible only through APIs, which limits transparency and local execution, or they are distributed openly, which deprives owners of monetization and control opportunities. We present OML (Open-access, Monetizable, and Loyal AI Model Serving), a primitive that enables a new distribution paradigm in which models can be freely distributed for local use while retaining cryptographically secured usage permissions.
------------------------------------------------------------
In short, OML is a thing that uses fingerprints. When looking at the open source code, you may not notice it right away, but OML assigns its own code to the model, and if you suspect that your model has been stolen, you can call up this 32-digit code, and if it matches, the copyright is yours! And yes, OML cannot be bypassed by combining models or tampering with them in any way. This code remains in the model anyway. By the way, I wrote in a comment when I read
@LeaderX_btc thread about OML that ,it reminded me a lot of a car's VIN code. If a thief steals a car, the police can simply look at the VIN code, which is well hidden in the car.
------------------------------------------------------------Also, OML introduces a control plane alongside the data plane, which is responsible for inference.
The control plane manages keys, verifies policies, and supports attestations. Each model launch is accompanied by a signed execution manifest (indicating who, what, and where) and an audit log that can only be appended to.
In OML, this allows you to:
📜Control who runs the model and with what permissions
📜Check compliance with policies (e.g., prohibiting certain types of requests)
📜Keep an audit log
📜Verify the origin of the model through fingerprints
Imagine a restaurant🐔
The kitchen is a data plane. This is where dishes are prepared (model inference). Everything happens quickly, according to recipes, and the result is a finished dish.
The restaurant manager and the ordering system are the management plane. It:
Checks who placed the order (authorization)
Takes into account which dishes are allowed (policies)
Keeps a log of orders (audit)
Can confirm that the dish was prepared in this particular restaurant (fingerprints)
The kitchen can operate without a manager, but then anyone can come in, take a recipe, change it, and no one will know where and by whom it was used. With a manager, everything is under control.
------------------------------------------------------------
The breakthrough of OML in the fact that its Open, Monetizable, and Loyal architecture provides developers and researchers with:
the ability to freely distribute model weights for scientific, educational, and public initiatives; the ability to track model usage and support sustainable economic mechanisms; the ability to enforce behavioral and policy rules established by the community or model owner.
In short; very cool stuff which allows you to publish open source code of the model without fear that someone will steal it.
#Sentient #OML #opensourcecode