Would LLMs like GitHub Copilot/Codex/ChatGPT write better code if they had access to some sort of data for what the purpose of each token in the given input has, analogous to how syntax highlighting helps humans write code?