Simon argues that using LLMs as a pair programming partner works well if you have test coverage. This gives you more guardrails around creating regressions. Makes sense to me in my experience with using LLMs with dev tasks.