The problem with training LLMs on internet data is that the data on the internet is finite. I had a strong feeling development would slow down eventually and it looks like that's happening.