TRAINING AI TO PAINT WITH CODE

Summary from the TLDR Design newsletter

A language model was trained with reinforcement learning to create watercolor images by writing editable p5.brush JavaScript rather than generating fixed image files, giving people direct control over the resulting artifact. Early training exposed problems with overly complex reward systems, while switching from absolute scores to pairwise comparisons against a hand-rated reference pool produced better images with dramatically shorter code. The experiment shows that applying reinforcement learning to creative work is fundamentally a design problem: subjective qualities like taste need carefully constructed reward structures that provide direction without forcing the model toward a single aesthetic.