Connects single-chain protein folding to a general-purpose language model, using its residue representations to drive a pretrained structure generator. Current results show weaker performance on longer chains and reduced scores on a general-knowledge test after protein training.