One thing I find particularly interesting is the separation between the reasoning model and the tool layer. Standards like MCP could make AI gd applications much more modular and interoperable.