Скачать или смотреть Breeding Python Programs (AI program synthesis + evolution)

Each image represents a single fully deterministic python program that generated this image. Using the CLIP model to provide evolutionary pressure the system is trying to generate programs that satisfy the currently selected goal as much as possible. Right now it is: “Photo of a real human face”.

While current generative AI models provide astonishing results, the interpretability of why these models work is usually almost impossible. With this approach we do not train the AI to generate the data, but rather programs that generate the data. That allows us to have a completely transparent code that could be analyzed in full detail.

Furthermore, it allows to create a simple automatic validation loop allowing text models to be trained on the best scoring programs increasing the code-generation capabilities of the initial model. Basically allowing to create an infinite improvement loop bounded only by the available compute.