moats in ai, built on shifting ground

I have been thinking about moat in building an AI company, and three independent insights from three different people make a lot of sense when put together: 1/ Friend 1: It is very difficult to unlock value and sustain it for long if the tech underneath is very volatile and the ground is constantly shifting. 2/ Friend 2: Ask yourself, are you happy or sad when the model underneath gets better? The answer shows whether you are doing the right things or not. 3/ Andrej: For founders, one important opportunity is to find domains that are valuable, verifiable, and undertrained by frontier labs. Build in domains where model outputs need continuous grounding, verification, and feedback loops that only your system accumulates over time.