Everybody training models on large amounts of lightly filtered internet text is partially distilling every other model that had its output posted verbatim to the internet.

And OpenAI probably distills anthropic, who would't?

It's all one big incestuous mess. In a couple of years we'll be talking about AI brainrot.