I post trained a model on Demoscenes

The author trained a language model on Dwitter.net's 140-character JavaScript demos to explore how incomplete programs could be completed in different ways, inspired by the vast number of possible continuations that exist within programming's finite rules. By studying how similar demos share code fragments while producing different visual effects, they discovered that an autoregressive model could suggest plausible completions for unfinished demos, allowing for creative exploration and mixing of different implementations. This approach combines linguistic productivity principles with demoscene aesthetics to help overcome creative blocks in generating visual effects.