Thanks for sharing that link. There are some really beautiful pieces in there! Natural, but with a strange edge. Regular enough to be pleasing, but unpredictable enough to be interesting.
It would be interesting to see this combined with simulation [1], which might itself need another neural network (or manually written "fixer") to interpret the ambiguous instructions.
The comment about 19th century patterns resonated. I don’t knit well, but I do tatting, and original vintage patterns are often woefully incomplete, there seemed to be this expectation that you knew the basics and could work out what was missing yourself. As The article points out, knowing the intent of the pattern usually helps resolve ambiguity
The naivety of these systems leads to output that we perceive as creative or playful. Personally, I've used simple Markov models to create texts or melodies that I then selectively recombine to write new music. The computer-created material is familiar but different enough from what I or a human might produce that it leads me to push my own creativity in new directions.
