NA
Over the past few years, I’ve been giving workshops about machine learning for art and creativity, and building an open-source toolkit for it called ml4a. As I learned about decentralized AI, I gradually became inspired by the idea of an autonomous artificial artist, a sovereign creative spirit who generates original art.
Baseline autonomous artificial artist: a trainable generative model whose data, code, curation, and governance are crowd-sourced from a decentralized network of actors. The behavior of the program emerges from its cumulative interactions with these actors.
To ensure the originality and uniqueness of sampled artworks, a model is required to be irreproducible and unforgeable. That is nobody — inside or outside of the AAA — is able to clone or retrain the same model, nor sample from it externally.
To meet the irreproducibility constraint, the model is trained blindly on crowd-sourced data which is never aggregated, instead remaining private to each of the individual contributors, leaving behind no easy way to recreate the same dataset a second time.
Uniqueness is secured by splitting the model into many pieces which are distributed throughout the network, and held together as a shared secret. To sample from the model, a query propagates through the entire network. Because no individual has the full model, it’s impossible to generate a sample any other way.
In contrast to depictions of AI as alien or oppositional to us, this project is a deeply humanistic one, conceiving of AI as a vehicle for the collective intelligence and creativity of people.