Georgi on why it's still hard to get great coding agent performance from local models: "Note that the main issues that people currently unknowingly face with local models mostly revolve around the harness and some intricacies around model chat templates and prompt construction"