Don’t Generate, Discriminate: A Proposal for Grounding Language Models to Real-World Environments