It feels like the pace of improvement for foundational models is already slowing down. The trajectory feels much more linear than a few years ago when we had massive, qualitative jumps like GPT-2 to GPT-3.