hmm i'm not totally convinced that an llm could be trained to produce good patchwork code without some significant natural language capability. it also seems interesting to compare/contrast these to the analogous claims with 'human' substituted for 'llm'