I came across your work a few weeks ago and was interested in the possibility in using dojo to implement provable model inference. The gist of it is Dojo is a provable game engine inspired by Rust. It is written in Cairo which means that we can generate a mathematical proof of computation using a zkSTARK. It creates an interesting dynamic where a player can train a neural network to drive the car, then run inference using the Dojo implementation, and a submit a proof of their score, without having to expose the model weights. It's mostly an experiment with this new affordance. Essentially, with provable computation it is no longer necessary to verify a clients computation using a trusted centralized server, which enables interesting new architectures and also fully private information.
I wanted to share here incase you found it interesting. Happy to share more if so!
Hey!
I came across your work a few weeks ago and was interested in the possibility in using dojo to implement provable model inference. The gist of it is Dojo is a provable game engine inspired by Rust. It is written in Cairo which means that we can generate a mathematical proof of computation using a zkSTARK. It creates an interesting dynamic where a player can train a neural network to drive the car, then run inference using the Dojo implementation, and a submit a proof of their score, without having to expose the model weights. It's mostly an experiment with this new affordance. Essentially, with provable computation it is no longer necessary to verify a clients computation using a trusted centralized server, which enables interesting new architectures and also fully private information.
I wanted to share here incase you found it interesting. Happy to share more if so!
Thanks for the original implementation!